本文最后更新于:2023年4月15日 晚上
本文记录探究 STPM 网络性能的实验过程、结果及分析。
概述

- 工作源自于论文《Student-Teacher Feature Pyramid Matching for Unsupervised Anomaly Detection》1
- 该方法需要预训练的教师网络,以及利用教师在下游OK数据上产生的特征蒸馏训练出学生网络;由两个网络产生的特征金字塔之间的差异作为评分函数,该得分表示数据中异常发生的概率。1
- 无监督学习在项目中有着重要地位:区别在于监督学习识别缺陷特征的模式,无监督学习使用OK数据进行训练得到模型,检出缺陷依赖的是缺陷图像在特征空间中的异常表达,这种不同的检测原理是该方法将成为监督学习重要补充的根源
|
监督学习 |
无监督异常检测 |
数据源 |
精确标注数据 |
二分类标注——OK部分数据 |
结果表达能力 |
数据子类别、位置、mask |
二分类 mask |
内因驱动 |
学习如何识别缺陷数据特征 |
依赖OK数据训练的系统在 NG 数据上的异常表示 |
- STPM 在无监督异常检测方法中对于
聚芯5000
项目的数据有特殊的优势:
- Ai 算法,自适应数据
- 不需要配置过多参数
- 相对于传统异常检测方法省去了大量调参过程,也具有一定泛化能力
- 接受底板、原图的组合输入并有效找出差异
- 接受不断变化的底板作为输入
- 可以有效检出监督学习的严重漏检数据
- 在数据质量高时过杀很少
已经在实验展开之前在长电
的数据中验证了可用性,本次实验旨在优化模型性能,对该方法的未解之处进行探究。
数据集
实验中使用的数据为 长电
缺陷检测设备输出的疑似NG图像,主要使用的有两组数据集
数据1
Part1 |
All / OK / NG |
Part2 |
All / OK / NG |
Part3 |
All / OK / NG |
MPS-MX3067R22-M1 |
88 / 5 / 83 |
MPS-SM3519ZR8-UBM |
1653 / 461 /1192 |
MPS-MX3067R22-UBM |
321 / 226 / 95 |
MPS-ST2769R19-UBM |
5083 / 1483 /3600 |
MPS-ST3031R1-UBM |
18751 / 11646 / 7105 |
MPS-ST9410R9-UBM |
4104 / 3508 / 596 |
MPS-ST3099R5-UBM |
4331 / 2468 / 1863 |
MPS-ST3405ZR5-M1 |
872 / 828 / 44 |
MPS-ST9414R12-UBM |
11526 / 10685 / 841 |
MPS-ST3620R11-UBM |
1433 / 268 / 1165 |
MPS-ST9410R9-M1 |
1673 / 330 / 1343 |
MPS-ST9430R18-UBM |
4308 / 3621 / 687 |
MPS-ST9430R18-M1 |
68 / 1 / 67 |
|
|
|
|
- 划分好数据集后,将OK、NG数据汇总,OK 数据按照 的比例分流到
train
, val
, test
数据,NG数据合入 test 数据中,得到三份数据,汇总得到第四份数据
|
OK_train |
OK_val |
OK_test / NG_test |
Part1 |
2957 |
422 |
846 / 6788 |
Part2 |
9285 |
1326 |
2654 / 9684 |
Part3 |
12628 |
1804 |
3608 / 2219 |
Total |
24870 |
3552 |
7108 / 18681 |
数据2
- 客户近期对
TI-CD3701C0AYCUR(MHC0AECP)-M1
底板数据的自动复检需求强烈,遂整理了第二份数据
|
OK_train |
OK_val |
OK_test / NG_test |
TI |
14184 |
2026 |
4054 / 7898 |
数据3
- TI 系列事实上有 9 套底板,在数据2 使用一段时间之后才整理出了新的 9 种底板的数据
 |
|
底板名称 |
TI-CD3701C0AYCUR(MHC0AECP)-BALLDROP |
简称 |
MH-Balldrop |
OK_train |
1402 |
OK_val |
285 |
OK_test / NG_test |
1638 / 107 |
 |
|
底板名称 |
TI-CD3701C0AYCUR(MHC0AECP)-M1 |
简称 |
MH-M1 |
OK_train |
11308 |
OK_val |
3577 |
OK_test / NG_test |
5381 / 995 |
 |
|
底板名称 |
TI-CD3701C0AYCUR(MHC0AECP)-UBM |
简称 |
MH-UBM |
OK_train |
639 |
OK_val |
121 |
OK_test / NG_test |
129 / 97 |
 |
|
底板名称 |
TI-CD3701C0AYCUR-BALLDROP |
简称 |
Balldrop |
OK_train |
649 |
OK_val |
129 |
OK_test / NG_test |
521 / 98 |
 |
|
底板名称 |
TI-CD3701C0AYCUR-M1 |
简称 |
M1 |
OK_train |
1127 |
OK_val |
315 |
OK_test / NG_test |
505 / 38 |
 |
|
底板名称 |
TI-CD3701C0AYCUR-UBM |
简称 |
UBM |
OK_train |
8581 |
OK_val |
756 |
OK_test / NG_test |
5226 / 434 |
 |
|
底板名称 |
TI-CD3701C0AYCUR(RFC0AECP)-BALLDROP |
简称 |
RF-Balldrop |
OK_train |
6176 |
OK_val |
1342 |
OK_test / NG_test |
3736 / 105 |
 |
|
底板名称 |
TI-CD3701C0AYCUR(RFC0AECP)-M1 |
简称 |
RF-M1 |
OK_train |
1093 |
OK_val |
94 |
OK_test / NG_test |
1526 / 30 |
 |
|
底板名称 |
TI-CD3701C0AYCUR(RFC0AECP)-UBM |
简称 |
RF-UBM |
OK_train |
2311 |
OK_val |
460 |
OK_test / NG_test |
813 / 89 |
探究目标
模型训练阶段
1
| STPM训练阶段基础参数Padding 训练数据量训练轮数batch size图像尺寸截图尺寸压缩尺寸数据抗噪能力数据增强去色操作常规增强数据策略损失函数定位精度主干网络网络类型网络交叉预训练模型迁移能力剪枝训练数据微调通道注意力空间注意力降采样训练学生独断降采样异构网络
|
模型推理阶段
1
| STPM推理阶段特征层级转化结果策略结果先验遮罩探索新的特征组合策略统计样本分值分布特征聚类分析
|
实验设置
探究内容 |
实验选项 |
batch size |
2, 6, 10 |
训练轮数 |
10, 20, 30, … , 100, 110 |
数据量 |
20% 40% 60% 80% 100% |
截图尺寸 |
600×600 400×400 256×256 |
缩放尺寸 |
600×600 400×400 256×256 |
抗噪性能 |
加噪 1% 2% 4% 8% |
去色预处理 |
性能测试、迁移能力鲁棒性测试 |
常规数据增强 |
翻转、旋转、gamma 变换 |
数据策略 |
教师-底板 学生-原图 vs 教师-原图 学生-底板 |
损失函数 |
L1, L2, Huber Loss, CosSim |
定位精度 |
4-8 pixel 1-2 pixel |
网络类型 |
Resnet 18 Resnet 34 Resnet 50 Resnet 101 WideResnet50 resnext50 densenet121 |
网络交叉 |
Resnet 18 × Resnet 34 Resnet 50 × WideResnet50 |
预训练 |
ImageNet , FasterRcnn coco Backbone Resnet 50, fackbook ssl, facebook swsl, ibn_a, ibn_b |
数据微调 |
用 TI 数据对 fackbook ssl 预训练 resnet50 模型finetune ,对比模型性能 |
模型迁移能力 |
Part1 Part2 Part3 Total 数据集相互测试 |
剪枝训练 |
使用网络不同层级特征进行训练,测试性能 |
Padding 训练 |
探究训练 STPM 网络时加入 Padding 对网络性能的影像 |
Hcsc 预训练网络 |
使用 Hcsc 方法,使用所有ADC数据(300k+)预训练模型 |
统计样本分值分布 |
统计模型预测样本分值在二分类数据上的分布情况 |
特征聚类分析 |
分析不同产品数据的特征分布,探究与迁移能力是否存在相关性 |
通道注意力 |
为通道特征加入注意力机制 |
空间注意力 |
|
降采样训练 |
|
学生独断 |
|
降采样异构网络 |
|
实验、评估方法
- 在实验设置、实验数据的实验组合中,控制变量进行模型训练
- 训练出的模型在所有推理策略上进行推理,得到若干组推理结果(图像异常的概率)
- 用人工标注的二分类 AP 值评估该训练、推理方法在该数据集上的性能
实验结果
结果说明
符号 |
含义 |
L1-L4 |
表示第一到第四层特征 |
CC |
Center Crop |
_L, _S |
Large, Small |
Gau |
高斯核 |
N |
不采用 mask |
Mul |
四层特征相乘 |
Sum |
四层特征相加 |
Batch Size
实验目的
- 探究不同 Batch Size 对 STPM 模型性能的影响
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
2, 6, 10 |
Epoch |
36 |
Init Learning Rate |
0.01 |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
BatchSize 2 |
0.866 |
0.882 |
0.881 |
0.895 |
0.829 |
0.907 |
0.924 |
0.924 |
0.931 |
0.860 |
0.902 |
0.909 |
0.904 |
0.910 |
0.856 |
0.892 |
0.895 |
0.896 |
0.899 |
0.865 |
0.911 |
0.922 |
0.890 |
0.915 |
0.862 |
0.910 |
0.918 |
0.913 |
0.921 |
0.859 |
BatchSize 6 |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
BatchSize 10 |
0.866 |
0.883 |
0.883 |
0.898 |
0.829 |
0.907 |
0.924 |
0.924 |
0.931 |
0.858 |
0.904 |
0.912 |
0.907 |
0.914 |
0.857 |
0.896 |
0.902 |
0.901 |
0.906 |
0.867 |
0.916 |
0.926 |
0.893 |
0.919 |
0.864 |
0.913 |
0.922 |
0.916 |
0.924 |
0.861 |
相关结论
- 在相同条件下,WideResnet50 在 BatchSize 为 2, 6, 10 下性能几乎没有变化
后期实验
- 根据以上结论,考虑到显卡显存与实验一致性,之后的实验 Batch Size 均设置为 6
训练轮数
实验目的
- 探究 训练轮数(epoch num) 对 STPM 模型性能的影响
实验设置
实验项目 |
实验设置 |
数据集 |
TI |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
10, 20, … , 100, 110 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
训练层级 |
layer4 |
实验结果
- 采用在
TI
数据 测试集 上的 二分类 AP 作为评判指标:
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
10 |
0.962 |
0.973 |
0.975 |
0.978 |
0.929 |
0.920 |
0.944 |
0.950 |
0.967 |
0.903 |
0.840 |
0.867 |
0.864 |
0.896 |
0.805 |
0.807 |
0.819 |
0.800 |
0.845 |
0.673 |
0.916 |
0.940 |
0.900 |
0.945 |
0.787 |
0.864 |
0.893 |
0.898 |
0.930 |
0.696 |
20 |
0.954 |
0.970 |
0.971 |
0.977 |
0.925 |
0.899 |
0.922 |
0.929 |
0.951 |
0.887 |
0.835 |
0.856 |
0.856 |
0.884 |
0.816 |
0.810 |
0.826 |
0.797 |
0.850 |
0.674 |
0.901 |
0.923 |
0.888 |
0.929 |
0.798 |
0.857 |
0.880 |
0.887 |
0.917 |
0.704 |
30 |
0.957 |
0.972 |
0.973 |
0.978 |
0.930 |
0.907 |
0.930 |
0.936 |
0.956 |
0.895 |
0.841 |
0.862 |
0.862 |
0.890 |
0.819 |
0.816 |
0.831 |
0.800 |
0.856 |
0.673 |
0.910 |
0.931 |
0.897 |
0.937 |
0.803 |
0.865 |
0.888 |
0.894 |
0.923 |
0.705 |
40 |
0.954 |
0.970 |
0.971 |
0.977 |
0.925 |
0.899 |
0.922 |
0.929 |
0.951 |
0.887 |
0.835 |
0.856 |
0.856 |
0.884 |
0.816 |
0.810 |
0.826 |
0.797 |
0.850 |
0.674 |
0.901 |
0.923 |
0.888 |
0.929 |
0.798 |
0.857 |
0.880 |
0.887 |
0.917 |
0.704 |
50 |
0.953 |
0.969 |
0.971 |
0.976 |
0.925 |
0.898 |
0.921 |
0.928 |
0.950 |
0.886 |
0.835 |
0.855 |
0.856 |
0.882 |
0.818 |
0.812 |
0.825 |
0.797 |
0.849 |
0.675 |
0.900 |
0.922 |
0.889 |
0.928 |
0.802 |
0.857 |
0.879 |
0.887 |
0.915 |
0.706 |
60 |
0.954 |
0.970 |
0.971 |
0.977 |
0.926 |
0.897 |
0.920 |
0.927 |
0.949 |
0.885 |
0.836 |
0.855 |
0.856 |
0.883 |
0.821 |
0.815 |
0.827 |
0.798 |
0.852 |
0.676 |
0.900 |
0.922 |
0.889 |
0.928 |
0.805 |
0.858 |
0.879 |
0.887 |
0.915 |
0.708 |
70 |
0.953 |
0.969 |
0.971 |
0.977 |
0.926 |
0.896 |
0.919 |
0.925 |
0.948 |
0.884 |
0.834 |
0.853 |
0.854 |
0.881 |
0.820 |
0.814 |
0.827 |
0.798 |
0.851 |
0.676 |
0.899 |
0.920 |
0.888 |
0.927 |
0.805 |
0.856 |
0.878 |
0.885 |
0.914 |
0.709 |
80 |
0.954 |
0.970 |
0.972 |
0.977 |
0.927 |
0.898 |
0.920 |
0.927 |
0.950 |
0.887 |
0.836 |
0.856 |
0.856 |
0.884 |
0.820 |
0.816 |
0.829 |
0.799 |
0.853 |
0.677 |
0.901 |
0.923 |
0.890 |
0.929 |
0.806 |
0.859 |
0.880 |
0.888 |
0.917 |
0.709 |
90 |
0.953 |
0.969 |
0.971 |
0.977 |
0.925 |
0.895 |
0.918 |
0.925 |
0.948 |
0.884 |
0.835 |
0.854 |
0.855 |
0.882 |
0.818 |
0.815 |
0.830 |
0.797 |
0.854 |
0.676 |
0.899 |
0.921 |
0.888 |
0.927 |
0.804 |
0.857 |
0.878 |
0.886 |
0.915 |
0.707 |
100 |
0.954 |
0.970 |
0.971 |
0.977 |
0.926 |
0.897 |
0.920 |
0.926 |
0.949 |
0.886 |
0.837 |
0.856 |
0.857 |
0.883 |
0.820 |
0.816 |
0.830 |
0.799 |
0.854 |
0.677 |
0.902 |
0.923 |
0.891 |
0.930 |
0.807 |
0.859 |
0.880 |
0.888 |
0.917 |
0.710 |
110 |
0.952 |
0.969 |
0.970 |
0.976 |
0.924 |
0.894 |
0.917 |
0.924 |
0.947 |
0.883 |
0.835 |
0.853 |
0.855 |
0.881 |
0.821 |
0.814 |
0.827 |
0.798 |
0.851 |
0.676 |
0.898 |
0.919 |
0.888 |
0.926 |
0.807 |
0.857 |
0.877 |
0.885 |
0.914 |
0.711 |
- 采用高斯 mask 的第二层特征作为获取结果的 map (L2_Gau_L)绘制折线图

相关结论
- 在
TI
的数据上,训练 10 轮(学习率0.01) 已经达到最优的性能了,可能略有过拟合
- 可能
TI
数据较为单一,不需要训练过多轮数,数据更加复杂的训练任务可以适当增加轮数,至少不会严重地削弱模型性能
数据量
实验目的
实验设置
实验项目 |
实验设置 |
数据集 |
Part1,Part2,Part3,Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
训练层级 |
layer4 |
训练集数据量百分比 |
100%,80%,60%,40%,20% |
数据量 |
OK_train |
OK_val |
OK_test / NG_test |
Part1 |
2957 |
422 |
846 / 6788 |
Part2 |
9285 |
1326 |
2654 / 9684 |
Part3 |
12628 |
1804 |
3608 / 2219 |
Total |
24870 |
3552 |
7108 / 18681 |
实验结果
NG AP |
100% |
80% |
60% |
40% |
20% |
P1-P1 |
0.969 |
0.969 |
0.968 |
0.967 |
0.965 |
P1-P2 |
0.912 |
0.909 |
0.904 |
0.900 |
0.891 |
P1-P3 |
0.662 |
0.657 |
0.652 |
0.645 |
0.633 |
P1-Total |
0.833 |
0.835 |
0.831 |
0.827 |
0.818 |
P2-P1 |
0.961 |
0.963 |
0.964 |
0.965 |
0.962 |
P2-P2 |
0.924 |
0.919 |
0.914 |
0.905 |
0.897 |
P2-P3 |
0.693 |
0.691 |
0.664 |
0.651 |
0.644 |
P2-Total |
0.808 |
0.818 |
0.822 |
0.828 |
0.826 |
P3-P1 |
0.939 |
0.938 |
0.937 |
0.938 |
0.939 |
P3-P2 |
0.855 |
0.854 |
0.852 |
0.851 |
0.857 |
P3-P3 |
0.695 |
0.693 |
0.694 |
0.699 |
0.704 |
P3-Total |
0.892 |
0.891 |
0.889 |
0.889 |
0.890 |
Total-Total |
0.932 |
0.933 |
0.931 |
0.928 |
0.922 |
- 采用高斯 mask 的第二层特征作为获取结果的 map (L2_Gau_L)绘制条形图

相关结论
- 大多数情况下,训练数据越多 → 模型性能越好,相反的情况极少,可以认为在训练中使用的数据量多多益善
Part3
作为训练数据时,减少数据量几乎没有造成性能下降,可能的原因是 Part3
训练数据量充足(12628组),即便20%的训练数据(2526组)已经足够训练 STPM 网络
Part1,2
作为训练数据时,数据量变小最多影响性能3%,没有对模型性能有绝对的影响
Part1
作为训练数据时,最小的训练数据量为 592 组数据,即 600 组左右的数据就可以训练出基本可用的模型
截图尺寸
实验目的
- 探究 输入数据截图尺寸对 STPM 模型性能的影响
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
截图尺寸 |
600×600,400×400, 256×256 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
256×256 |
0.905 |
0.914 |
0.919 |
0.926 |
0.886 |
0.933 |
0.939 |
0.941 |
0.950 |
0.921 |
0.930 |
0.934 |
0.935 |
0.941 |
0.916 |
0.929 |
0.924 |
0.929 |
0.925 |
0.913 |
0.942 |
0.945 |
0.939 |
0.948 |
0.925 |
0.940 |
0.939 |
0.943 |
0.948 |
0.920 |
400×400 |
0.891 |
0.908 |
0.906 |
0.916 |
0.869 |
0.927 |
0.931 |
0.933 |
0.940 |
0.900 |
0.922 |
0.925 |
0.924 |
0.930 |
0.895 |
0.911 |
0.913 |
0.916 |
0.914 |
0.897 |
0.933 |
0.936 |
0.924 |
0.936 |
0.904 |
0.929 |
0.931 |
0.932 |
0.937 |
0.899 |
600×600 |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制折线图

相关结论
- 实验结果表明center crop 可以提升模型在当前数据集上的性能;
- 同时说明当前数据具有缺陷集中在中间的先验知识,不能作为 STPM 模型的相关结论。
缩放尺寸
实验目的
- 探究 resize 尺寸对 STPM 模型性能的影响
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
Resnet18, Resnet34, Resnet50, WideResnet50 |
缩放尺寸 |
256×256, 400×400, 600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Resnet18_256 |
0.853 |
0.855 |
0.855 |
0.859 |
0.828 |
0.887 |
0.885 |
0.887 |
0.885 |
0.870 |
0.883 |
0.886 |
0.883 |
0.886 |
0.869 |
0.906 |
0.910 |
0.909 |
0.910 |
0.904 |
0.905 |
0.909 |
0.898 |
0.906 |
0.886 |
0.909 |
0.915 |
0.914 |
0.918 |
0.898 |
Resnet18_400 |
0.853 |
0.854 |
0.855 |
0.859 |
0.819 |
0.891 |
0.894 |
0.893 |
0.893 |
0.861 |
0.887 |
0.892 |
0.886 |
0.890 |
0.863 |
0.892 |
0.899 |
0.897 |
0.903 |
0.883 |
0.902 |
0.908 |
0.891 |
0.901 |
0.874 |
0.900 |
0.908 |
0.905 |
0.910 |
0.882 |
Resnet18_600 |
0.845 |
0.864 |
0.864 |
0.874 |
0.809 |
0.896 |
0.906 |
0.904 |
0.908 |
0.854 |
0.894 |
0.899 |
0.893 |
0.899 |
0.853 |
0.899 |
0.904 |
0.905 |
0.907 |
0.885 |
0.907 |
0.915 |
0.892 |
0.910 |
0.868 |
0.905 |
0.912 |
0.910 |
0.915 |
0.881 |
Resnet34_256 |
0.857 |
0.861 |
0.860 |
0.862 |
0.829 |
0.883 |
0.884 |
0.882 |
0.885 |
0.860 |
0.890 |
0.892 |
0.891 |
0.893 |
0.878 |
0.903 |
0.909 |
0.905 |
0.910 |
0.897 |
0.906 |
0.911 |
0.896 |
0.905 |
0.883 |
0.906 |
0.914 |
0.910 |
0.917 |
0.893 |
Resnet34_400 |
0.855 |
0.860 |
0.860 |
0.862 |
0.816 |
0.891 |
0.899 |
0.895 |
0.896 |
0.858 |
0.893 |
0.896 |
0.893 |
0.897 |
0.874 |
0.902 |
0.903 |
0.904 |
0.904 |
0.885 |
0.908 |
0.912 |
0.894 |
0.905 |
0.875 |
0.908 |
0.911 |
0.909 |
0.911 |
0.882 |
Resnet34_600 |
0.847 |
0.861 |
0.860 |
0.869 |
0.811 |
0.896 |
0.912 |
0.909 |
0.915 |
0.854 |
0.893 |
0.902 |
0.897 |
0.904 |
0.859 |
0.903 |
0.907 |
0.905 |
0.910 |
0.875 |
0.910 |
0.919 |
0.894 |
0.913 |
0.868 |
0.907 |
0.914 |
0.910 |
0.917 |
0.873 |
Resnet50_256 |
0.865 |
0.866 |
0.867 |
0.866 |
0.826 |
0.893 |
0.897 |
0.894 |
0.896 |
0.857 |
0.889 |
0.892 |
0.889 |
0.893 |
0.867 |
0.890 |
0.895 |
0.895 |
0.897 |
0.879 |
0.903 |
0.906 |
0.889 |
0.901 |
0.870 |
0.898 |
0.902 |
0.898 |
0.902 |
0.871 |
Resnet50_400 |
0.864 |
0.874 |
0.873 |
0.877 |
0.822 |
0.902 |
0.911 |
0.908 |
0.911 |
0.854 |
0.894 |
0.899 |
0.892 |
0.899 |
0.857 |
0.894 |
0.896 |
0.896 |
0.898 |
0.874 |
0.904 |
0.911 |
0.888 |
0.903 |
0.866 |
0.901 |
0.907 |
0.900 |
0.908 |
0.863 |
Resnet50_600 |
0.858 |
0.868 |
0.871 |
0.880 |
0.819 |
0.892 |
0.912 |
0.909 |
0.917 |
0.852 |
0.895 |
0.904 |
0.897 |
0.905 |
0.847 |
0.891 |
0.895 |
0.893 |
0.897 |
0.870 |
0.903 |
0.913 |
0.889 |
0.905 |
0.862 |
0.900 |
0.909 |
0.902 |
0.910 |
0.859 |
WideResnet50_256 |
0.870 |
0.881 |
0.878 |
0.884 |
0.833 |
0.888 |
0.890 |
0.889 |
0.892 |
0.852 |
0.889 |
0.892 |
0.887 |
0.890 |
0.865 |
0.908 |
0.909 |
0.910 |
0.907 |
0.897 |
0.906 |
0.911 |
0.888 |
0.903 |
0.869 |
0.907 |
0.911 |
0.906 |
0.909 |
0.875 |
WideResnet50_400 |
0.870 |
0.890 |
0.886 |
0.896 |
0.831 |
0.899 |
0.907 |
0.905 |
0.908 |
0.856 |
0.895 |
0.901 |
0.896 |
0.900 |
0.864 |
0.899 |
0.894 |
0.903 |
0.896 |
0.883 |
0.908 |
0.914 |
0.890 |
0.906 |
0.868 |
0.908 |
0.912 |
0.907 |
0.911 |
0.871 |
WideResnet50_600 |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
- 采用高斯mask的第二层特征作为获取结果的map (L2_Gau_L)绘制折线图

相关结论
- 不同模型在不同尺寸上得到了相同的性能变化趋势
- 原图为 ,resize 变小尺寸会使得 STPM 性能下降
抗噪性能
实验目的
-
探究 噪声数据 对 STPM 模型性能的影响
-
噪声数据为有意错误翻转的图像:


实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss, Huber Loss |
噪声数据比例 |
0%, 1%, 2%, 4%, 8% |
- 实验结果
- 采用在 TI 数据 测试集 上的 二分类 AP 作为评判指标:
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Noise Data 1% |
0.869 |
0.888 |
0.887 |
0.902 |
0.832 |
0.910 |
0.926 |
0.926 |
0.933 |
0.862 |
0.906 |
0.913 |
0.908 |
0.914 |
0.860 |
0.897 |
0.900 |
0.901 |
0.904 |
0.870 |
0.917 |
0.928 |
0.896 |
0.921 |
0.867 |
0.915 |
0.922 |
0.917 |
0.924 |
0.864 |
Noise Data 2% |
0.871 |
0.887 |
0.887 |
0.901 |
0.835 |
0.910 |
0.924 |
0.924 |
0.931 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.897 |
0.900 |
0.902 |
0.904 |
0.872 |
0.917 |
0.927 |
0.898 |
0.921 |
0.869 |
0.914 |
0.922 |
0.917 |
0.925 |
0.866 |
Noise Data 4% |
0.868 |
0.886 |
0.885 |
0.900 |
0.833 |
0.909 |
0.924 |
0.924 |
0.930 |
0.862 |
0.905 |
0.912 |
0.907 |
0.914 |
0.861 |
0.897 |
0.901 |
0.902 |
0.904 |
0.871 |
0.917 |
0.926 |
0.898 |
0.920 |
0.868 |
0.914 |
0.921 |
0.916 |
0.924 |
0.866 |
Noise Data 8% |
0.856 |
0.873 |
0.873 |
0.888 |
0.821 |
0.903 |
0.919 |
0.918 |
0.924 |
0.853 |
0.902 |
0.910 |
0.905 |
0.911 |
0.856 |
0.895 |
0.900 |
0.900 |
0.903 |
0.870 |
0.912 |
0.923 |
0.892 |
0.916 |
0.862 |
0.910 |
0.918 |
0.913 |
0.920 |
0.862 |
Noise Data 8% HuberLoss |
0.866 |
0.883 |
0.883 |
0.897 |
0.831 |
0.908 |
0.922 |
0.922 |
0.928 |
0.860 |
0.905 |
0.912 |
0.908 |
0.914 |
0.859 |
0.896 |
0.900 |
0.902 |
0.904 |
0.879 |
0.916 |
0.925 |
0.897 |
0.919 |
0.867 |
0.913 |
0.920 |
0.915 |
0.922 |
0.867 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制折线图

相关结论
- STPM 训练过程中允许存在部分噪声(完全不匹配),噪声数据4%以下不会显著降低模型性能
- 8% 左右的噪声数据已经影响了模型性能,浅层特征(1-2层)性能下降,深层特征(3-4层)性能变化不大
- HuberLoss 应用在带噪数据上可以一定程度上抵抗噪声数据对模型带来的不良影响
去色预处理 —— 性能测试
实验由来
- 原图为彩色图像,而底板为灰度图像,此种差异是否会影响STPM 网络模型的性能?
实验目的
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
预处理操作 |
ToGray |
实验结果
- 采用在
Total
数据 测试集 上的 二分类 AP 作为评判指标:
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
BaseLine |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
ToGray |
0.859 |
0.880 |
0.878 |
0.894 |
0.821 |
0.908 |
0.925 |
0.924 |
0.931 |
0.858 |
0.904 |
0.910 |
0.906 |
0.911 |
0.854 |
0.894 |
0.898 |
0.899 |
0.902 |
0.866 |
0.915 |
0.924 |
0.892 |
0.918 |
0.861 |
0.912 |
0.919 |
0.914 |
0.922 |
0.859 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制柱状图

相关结论
去色预处理 —— 迁移能力测试
实验由来
- 不同产品图像的颜色大相径庭,是否去除这一差异会增强模型的迁移能力?
实验目的
- 探究 去色预处理 对 STPM 模型数据迁移能力的影响
实验设置
实验项目 |
实验设置 |
数据集 |
Part1, Part2, Part3, Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
预处理操作 |
ToGray |
实验结果
- 采用在
Part1, Part2, Part3, Total
数据 测试集 上的 二分类 AP 作为评判指标:
BaseLine NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Train Part1-Test Part1 |
0.916 |
0.937 |
0.937 |
0.949 |
0.902 |
0.960 |
0.969 |
0.969 |
0.976 |
0.936 |
0.959 |
0.965 |
0.963 |
0.967 |
0.936 |
0.962 |
0.965 |
0.962 |
0.967 |
0.946 |
0.966 |
0.970 |
0.952 |
0.968 |
0.938 |
0.965 |
0.970 |
0.969 |
0.973 |
0.941 |
Train Part1-Test Part2 |
0.805 |
0.795 |
0.801 |
0.813 |
0.809 |
0.894 |
0.909 |
0.912 |
0.917 |
0.882 |
0.835 |
0.849 |
0.844 |
0.858 |
0.843 |
0.797 |
0.847 |
0.839 |
0.862 |
0.824 |
0.858 |
0.876 |
0.869 |
0.877 |
0.863 |
0.832 |
0.861 |
0.854 |
0.876 |
0.843 |
Train Part1-Test Part3 |
0.471 |
0.477 |
0.484 |
0.504 |
0.421 |
0.646 |
0.655 |
0.662 |
0.675 |
0.574 |
0.620 |
0.626 |
0.636 |
0.650 |
0.546 |
0.525 |
0.548 |
0.553 |
0.541 |
0.459 |
0.633 |
0.631 |
0.624 |
0.639 |
0.534 |
0.611 |
0.614 |
0.624 |
0.626 |
0.506 |
Train Part1-Test Total |
0.748 |
0.747 |
0.751 |
0.762 |
0.742 |
0.811 |
0.828 |
0.833 |
0.844 |
0.784 |
0.752 |
0.759 |
0.760 |
0.765 |
0.739 |
0.712 |
0.731 |
0.731 |
0.733 |
0.714 |
0.750 |
0.757 |
0.754 |
0.759 |
0.738 |
0.738 |
0.753 |
0.753 |
0.759 |
0.719 |
Train Part2-Test Part1 |
0.918 |
0.939 |
0.939 |
0.949 |
0.890 |
0.950 |
0.962 |
0.961 |
0.968 |
0.920 |
0.948 |
0.947 |
0.951 |
0.948 |
0.919 |
0.904 |
0.909 |
0.909 |
0.923 |
0.905 |
0.945 |
0.951 |
0.943 |
0.952 |
0.914 |
0.932 |
0.935 |
0.938 |
0.945 |
0.921 |
Train Part2-Test Part2 |
0.864 |
0.867 |
0.867 |
0.886 |
0.871 |
0.898 |
0.924 |
0.924 |
0.935 |
0.890 |
0.889 |
0.905 |
0.901 |
0.909 |
0.887 |
0.900 |
0.904 |
0.911 |
0.908 |
0.908 |
0.906 |
0.920 |
0.917 |
0.918 |
0.905 |
0.901 |
0.914 |
0.913 |
0.923 |
0.901 |
Train Part2-Test Part3 |
0.578 |
0.620 |
0.634 |
0.661 |
0.449 |
0.672 |
0.686 |
0.693 |
0.686 |
0.617 |
0.579 |
0.597 |
0.604 |
0.615 |
0.532 |
0.400 |
0.406 |
0.415 |
0.432 |
0.383 |
0.605 |
0.624 |
0.614 |
0.635 |
0.509 |
0.531 |
0.547 |
0.558 |
0.574 |
0.453 |
Train Part2-Test Total |
0.780 |
0.799 |
0.805 |
0.821 |
0.765 |
0.775 |
0.803 |
0.808 |
0.828 |
0.754 |
0.727 |
0.741 |
0.743 |
0.752 |
0.716 |
0.679 |
0.675 |
0.681 |
0.682 |
0.675 |
0.722 |
0.734 |
0.731 |
0.739 |
0.709 |
0.702 |
0.708 |
0.713 |
0.719 |
0.689 |
Train Part3-Test Part1 |
0.902 |
0.917 |
0.921 |
0.925 |
0.897 |
0.937 |
0.936 |
0.939 |
0.942 |
0.934 |
0.932 |
0.927 |
0.932 |
0.933 |
0.929 |
0.876 |
0.876 |
0.879 |
0.883 |
0.880 |
0.924 |
0.925 |
0.928 |
0.931 |
0.921 |
0.911 |
0.911 |
0.916 |
0.921 |
0.910 |
Train Part3-Test Part2 |
0.822 |
0.833 |
0.842 |
0.850 |
0.824 |
0.834 |
0.851 |
0.855 |
0.863 |
0.823 |
0.812 |
0.814 |
0.817 |
0.820 |
0.818 |
0.779 |
0.787 |
0.781 |
0.783 |
0.780 |
0.809 |
0.822 |
0.821 |
0.827 |
0.814 |
0.785 |
0.803 |
0.803 |
0.811 |
0.800 |
Train Part3-Test Part3 |
0.685 |
0.745 |
0.743 |
0.772 |
0.528 |
0.695 |
0.714 |
0.707 |
0.719 |
0.612 |
0.723 |
0.737 |
0.717 |
0.736 |
0.606 |
0.774 |
0.780 |
0.759 |
0.781 |
0.573 |
0.762 |
0.773 |
0.685 |
0.749 |
0.602 |
0.769 |
0.769 |
0.752 |
0.766 |
0.593 |
Train Part3-Test Total |
0.853 |
0.866 |
0.873 |
0.880 |
0.826 |
0.882 |
0.890 |
0.892 |
0.897 |
0.863 |
0.864 |
0.863 |
0.867 |
0.869 |
0.856 |
0.816 |
0.817 |
0.815 |
0.818 |
0.814 |
0.857 |
0.863 |
0.863 |
0.868 |
0.850 |
0.836 |
0.845 |
0.848 |
0.855 |
0.836 |
Gray NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Train Part1-Test Part1 |
0.920 |
0.937 |
0.938 |
0.949 |
0.907 |
0.959 |
0.969 |
0.969 |
0.975 |
0.939 |
0.957 |
0.964 |
0.962 |
0.967 |
0.935 |
0.962 |
0.965 |
0.962 |
0.968 |
0.946 |
0.966 |
0.970 |
0.954 |
0.969 |
0.939 |
0.964 |
0.970 |
0.968 |
0.973 |
0.941 |
Train Part1-Test Part2 |
0.784 |
0.781 |
0.785 |
0.796 |
0.787 |
0.894 |
0.910 |
0.911 |
0.918 |
0.879 |
0.840 |
0.855 |
0.849 |
0.862 |
0.848 |
0.802 |
0.851 |
0.845 |
0.868 |
0.826 |
0.862 |
0.880 |
0.873 |
0.879 |
0.864 |
0.832 |
0.865 |
0.857 |
0.881 |
0.844 |
Train Part1-Test Part3 |
0.454 |
0.457 |
0.465 |
0.484 |
0.414 |
0.638 |
0.648 |
0.657 |
0.669 |
0.565 |
0.601 |
0.614 |
0.622 |
0.637 |
0.535 |
0.515 |
0.529 |
0.532 |
0.522 |
0.442 |
0.606 |
0.614 |
0.598 |
0.621 |
0.516 |
0.584 |
0.591 |
0.597 |
0.604 |
0.490 |
Train Part1-Test Total |
0.736 |
0.736 |
0.740 |
0.750 |
0.730 |
0.826 |
0.844 |
0.849 |
0.859 |
0.794 |
0.758 |
0.766 |
0.767 |
0.772 |
0.745 |
0.716 |
0.729 |
0.731 |
0.732 |
0.718 |
0.754 |
0.761 |
0.758 |
0.762 |
0.740 |
0.740 |
0.754 |
0.754 |
0.761 |
0.722 |
Train Part2-Test Part1 |
0.928 |
0.943 |
0.944 |
0.952 |
0.904 |
0.948 |
0.962 |
0.961 |
0.970 |
0.916 |
0.944 |
0.947 |
0.948 |
0.950 |
0.915 |
0.907 |
0.914 |
0.912 |
0.927 |
0.906 |
0.944 |
0.954 |
0.942 |
0.953 |
0.911 |
0.932 |
0.938 |
0.938 |
0.948 |
0.918 |
Train Part2-Test Part2 |
0.829 |
0.830 |
0.827 |
0.844 |
0.829 |
0.885 |
0.909 |
0.908 |
0.919 |
0.879 |
0.884 |
0.898 |
0.895 |
0.902 |
0.884 |
0.898 |
0.902 |
0.909 |
0.906 |
0.907 |
0.899 |
0.911 |
0.912 |
0.909 |
0.903 |
0.896 |
0.907 |
0.906 |
0.915 |
0.899 |
Train Part2-Test Part3 |
0.474 |
0.485 |
0.497 |
0.526 |
0.419 |
0.589 |
0.615 |
0.624 |
0.640 |
0.530 |
0.549 |
0.568 |
0.572 |
0.593 |
0.496 |
0.380 |
0.392 |
0.397 |
0.414 |
0.368 |
0.537 |
0.561 |
0.548 |
0.572 |
0.457 |
0.496 |
0.517 |
0.527 |
0.552 |
0.423 |
Train Part2-Test Total |
0.747 |
0.754 |
0.757 |
0.770 |
0.736 |
0.768 |
0.798 |
0.804 |
0.826 |
0.744 |
0.726 |
0.739 |
0.740 |
0.749 |
0.712 |
0.682 |
0.679 |
0.684 |
0.684 |
0.676 |
0.710 |
0.725 |
0.720 |
0.730 |
0.697 |
0.699 |
0.706 |
0.710 |
0.718 |
0.684 |
Train Part3-Test Part1 |
0.898 |
0.903 |
0.907 |
0.914 |
0.901 |
0.929 |
0.926 |
0.929 |
0.930 |
0.923 |
0.926 |
0.923 |
0.927 |
0.928 |
0.920 |
0.876 |
0.876 |
0.878 |
0.880 |
0.877 |
0.916 |
0.919 |
0.919 |
0.922 |
0.908 |
0.904 |
0.908 |
0.912 |
0.916 |
0.896 |
Train Part3-Test Part2 |
0.787 |
0.790 |
0.793 |
0.795 |
0.787 |
0.811 |
0.811 |
0.812 |
0.812 |
0.797 |
0.767 |
0.767 |
0.768 |
0.771 |
0.767 |
0.745 |
0.756 |
0.745 |
0.749 |
0.760 |
0.771 |
0.772 |
0.774 |
0.774 |
0.775 |
0.749 |
0.758 |
0.756 |
0.763 |
0.762 |
Train Part3-Test Part3 |
0.647 |
0.716 |
0.713 |
0.740 |
0.504 |
0.690 |
0.708 |
0.700 |
0.705 |
0.604 |
0.716 |
0.725 |
0.709 |
0.721 |
0.593 |
0.768 |
0.780 |
0.756 |
0.781 |
0.553 |
0.754 |
0.761 |
0.678 |
0.738 |
0.589 |
0.765 |
0.760 |
0.744 |
0.752 |
0.576 |
Train Part3-Test Total |
0.830 |
0.838 |
0.842 |
0.848 |
0.801 |
0.867 |
0.866 |
0.868 |
0.868 |
0.844 |
0.838 |
0.839 |
0.841 |
0.844 |
0.826 |
0.797 |
0.801 |
0.796 |
0.800 |
0.801 |
0.834 |
0.836 |
0.834 |
0.838 |
0.824 |
0.816 |
0.823 |
0.824 |
0.830 |
0.813 |
相关结论
- 去色操作几乎在所有对比实验中的性能都有下降,因此去色会降低模型的数据迁移能力
- 在
Total
数据集上,彩色不是影响模型迁移性能的关键因素
常规数据增强
实验目的
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
数据增强 |
Random Flip Rotate,gamma 变换 |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Base |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
Random Flip Rotate |
0.860 |
0.878 |
0.877 |
0.890 |
0.825 |
0.906 |
0.922 |
0.923 |
0.929 |
0.855 |
0.900 |
0.904 |
0.901 |
0.904 |
0.850 |
0.892 |
0.899 |
0.898 |
0.902 |
0.858 |
0.911 |
0.922 |
0.889 |
0.916 |
0.856 |
0.910 |
0.917 |
0.913 |
0.918 |
0.853 |
Random Flip Rotate gamma |
0.863 |
0.879 |
0.879 |
0.892 |
0.826 |
0.907 |
0.924 |
0.924 |
0.929 |
0.856 |
0.901 |
0.906 |
0.902 |
0.905 |
0.850 |
0.892 |
0.899 |
0.898 |
0.904 |
0.859 |
0.913 |
0.923 |
0.891 |
0.917 |
0.856 |
0.910 |
0.918 |
0.914 |
0.919 |
0.853 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制折线图

相关结论
- 数据增强没有在同源的测试集上看到性能收益
- 可能由于数据成像稳定,原图与测试图像同源程度高,增强反而降低了模型性能,潜在的泛化性可能得到了提升
网络类型 测试一
实验目的
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
Resnet18, Resnet34, Resnet50, WideResnet50, resnext_50, densenet121, Swin |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
实验结果
常规卷积模型采用 尺寸
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Resnet18 |
0.845 |
0.864 |
0.864 |
0.874 |
0.809 |
0.896 |
0.906 |
0.904 |
0.908 |
0.854 |
0.894 |
0.899 |
0.893 |
0.899 |
0.853 |
0.899 |
0.904 |
0.905 |
0.907 |
0.885 |
0.907 |
0.915 |
0.892 |
0.910 |
0.868 |
0.905 |
0.912 |
0.910 |
0.915 |
0.881 |
Resnet34 |
0.847 |
0.861 |
0.860 |
0.869 |
0.811 |
0.896 |
0.912 |
0.909 |
0.915 |
0.854 |
0.893 |
0.902 |
0.897 |
0.904 |
0.859 |
0.903 |
0.907 |
0.905 |
0.910 |
0.875 |
0.910 |
0.919 |
0.894 |
0.913 |
0.868 |
0.907 |
0.914 |
0.910 |
0.917 |
0.873 |
Resnet50 |
0.858 |
0.868 |
0.871 |
0.880 |
0.819 |
0.892 |
0.912 |
0.909 |
0.917 |
0.852 |
0.895 |
0.904 |
0.897 |
0.905 |
0.847 |
0.891 |
0.895 |
0.893 |
0.897 |
0.870 |
0.903 |
0.913 |
0.889 |
0.905 |
0.862 |
0.900 |
0.909 |
0.902 |
0.910 |
0.859 |
WideResnet50 |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
resnext_50 |
0.852 |
0.879 |
0.878 |
0.891 |
0.809 |
0.900 |
0.915 |
0.914 |
0.918 |
0.851 |
0.899 |
0.902 |
0.899 |
0.902 |
0.854 |
0.901 |
0.903 |
0.903 |
0.904 |
0.881 |
0.913 |
0.919 |
0.896 |
0.915 |
0.868 |
0.908 |
0.913 |
0.910 |
0.913 |
0.877 |
densenet121 |
0.781 |
0.763 |
0.770 |
0.766 |
0.790 |
0.898 |
0.916 |
0.911 |
0.918 |
0.845 |
0.897 |
0.902 |
0.899 |
0.902 |
0.863 |
0.902 |
0.904 |
0.907 |
0.908 |
0.890 |
0.908 |
0.915 |
0.890 |
0.907 |
0.866 |
0.906 |
0.911 |
0.910 |
0.914 |
0.887 |
Swin Backbone 预训练时 ,对比 Resnet 50 的模型:
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
swin_base_384 |
0.813524 |
0.807814 |
0.814646 |
0.812689 |
0.809406 |
0.87153 |
0.883874 |
0.880479 |
0.887091 |
0.838493 |
0.875963 |
0.879824 |
0.874542 |
0.878654 |
0.855604 |
0.823823 |
0.818704 |
0.821138 |
0.818358 |
0.828502 |
0.895297 |
0.901938 |
0.891311 |
0.900453 |
0.87737 |
0.840747 |
0.841286 |
0.840095 |
0.844756 |
0.839377 |
Resnet50_400 |
0.863514 |
0.873812 |
0.872901 |
0.877489 |
0.821922 |
0.902308 |
0.911151 |
0.908467 |
0.911027 |
0.853546 |
0.8937 |
0.898684 |
0.892007 |
0.899019 |
0.857087 |
0.894162 |
0.895793 |
0.896131 |
0.897575 |
0.87415 |
0.90424 |
0.911145 |
0.887676 |
0.903215 |
0.865576 |
0.900666 |
0.906703 |
0.900253 |
0.9078 |
0.863266 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制折线图

相关结论
- 在相同条件下,WideResnet50 在各个层次特征中表现最佳
- Swin 表现不如 Resnet50
后期实验
- 根据以上结论,之后的实验大多数情况下在 WideResnet50 下进行
网络类型 测试二
实验目的
- 使用
数据3
作为训练、测试数据,再次测试不同网络类型的性能
实验设置
实验项目 |
实验设置 |
数据集 |
数据3 —— 9套TI底板数据 |
模型 |
Wide-Resnet50, Resnet50, Resnet18 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
实验结果
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 作为评判指标:
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Wide Resnet50 |
0.919 |
0.735 |
0.859 |
0.454 |
0.904 |
0.898 |
0.969 |
0.969 |
0.908 |
Resnet50 |
0.913 |
0.769 |
0.913 |
0.400 |
0.972 |
0.940 |
0.953 |
0.976 |
0.954 |
Resnet18 |
0.850 |
0.679 |
0.899 |
0.368 |
0.931 |
0.898 |
0.922 |
0.938 |
0.910 |
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的AP拐点NG得分作为阈值指标:
score |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Wide Resnet50 |
1.916318 |
1.55782 |
1.562743 |
1.743724 |
1.882993 |
1.787696 |
1.576425 |
1.929622 |
1.506365 |
Resnet50 |
2.015119 |
1.773723 |
1.749128 |
2.012612 |
1.869097 |
1.881239 |
2.011937 |
2.061737 |
1.550954 |
Resnet18 |
2.479775 |
2.103164 |
2.165806 |
2.697112 |
2.411225 |
2.367473 |
2.44851 |
2.420266 |
2.013886 |
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 绘制折线图:

相关结论
- 在
数据3
上 Resnet50 的性能更佳
- 可能数据种类较多的情况下需要更大的模型,数据多样性小的话小一些的模型更适用
- 在不同数据上需要不同的 backbone 网络,可以多实验几个模型
预训练模型
实验目的
- 探究 不同数据预训练 Resnet50 对 STPM 模型性能的影响
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
Resnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet,Coco,YFCC100M, ibn-a, ibn-b, ImageNet1K … |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
base |
0.866 |
0.879 |
0.882 |
0.893 |
0.825 |
0.897 |
0.916 |
0.913 |
0.922 |
0.855 |
0.899 |
0.909 |
0.902 |
0.911 |
0.844 |
0.890 |
0.899 |
0.871 |
0.902 |
0.790 |
0.906 |
0.920 |
0.887 |
0.909 |
0.859 |
0.902 |
0.915 |
0.906 |
0.917 |
0.837 |
ibn-a |
0.867 |
0.885 |
0.885 |
0.901 |
0.832 |
0.903 |
0.919 |
0.919 |
0.925 |
0.857 |
0.894 |
0.904 |
0.899 |
0.908 |
0.858 |
0.865 |
0.883 |
0.841 |
0.880 |
0.802 |
0.910 |
0.920 |
0.885 |
0.911 |
0.861 |
0.881 |
0.900 |
0.866 |
0.904 |
0.820 |
ibn-b |
0.868 |
0.890 |
0.888 |
0.900 |
0.837 |
0.894 |
0.912 |
0.913 |
0.922 |
0.861 |
0.893 |
0.905 |
0.897 |
0.909 |
0.842 |
0.894 |
0.895 |
0.861 |
0.897 |
0.816 |
0.907 |
0.917 |
0.887 |
0.910 |
0.862 |
0.902 |
0.909 |
0.882 |
0.913 |
0.822 |
ssl |
0.860 |
0.883 |
0.883 |
0.896 |
0.827 |
0.897 |
0.916 |
0.914 |
0.923 |
0.856 |
0.897 |
0.905 |
0.900 |
0.909 |
0.857 |
0.903 |
0.905 |
0.893 |
0.909 |
0.835 |
0.907 |
0.918 |
0.890 |
0.910 |
0.865 |
0.904 |
0.912 |
0.907 |
0.915 |
0.852 |
swsl |
0.859 |
0.883 |
0.884 |
0.896 |
0.816 |
0.891 |
0.908 |
0.907 |
0.915 |
0.845 |
0.898 |
0.904 |
0.902 |
0.909 |
0.857 |
0.912 |
0.908 |
0.844 |
0.913 |
0.816 |
0.915 |
0.923 |
0.892 |
0.917 |
0.863 |
0.915 |
0.912 |
0.866 |
0.917 |
0.834 |
coco fpn |
0.885 |
0.892 |
0.889 |
0.893 |
0.867 |
0.902 |
0.898 |
0.904 |
0.901 |
0.895 |
0.901 |
0.894 |
0.902 |
0.899 |
0.901 |
0.900 |
0.894 |
0.900 |
0.898 |
0.900 |
0.908 |
0.905 |
0.906 |
0.910 |
0.903 |
0.904 |
0.898 |
0.904 |
0.902 |
0.904 |
coco |
0.776 |
0.777 |
0.779 |
0.787 |
0.764 |
0.878 |
0.896 |
0.895 |
0.899 |
0.846 |
0.896 |
0.901 |
0.898 |
0.901 |
0.874 |
0.906 |
0.907 |
0.906 |
0.908 |
0.885 |
0.909 |
0.916 |
0.900 |
0.912 |
0.881 |
0.907 |
0.911 |
0.908 |
0.911 |
0.883 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制折线图

相关结论
- 不同数据和训练方法在同样的网络结构(Resnet50)上应用到 STPM 有不同的表现,其中 SSL 方法训练的模型综合来看效果较好;
- ImageNet 训练的网络(base) 也是可用的;
- coco 两组实验是从 faster rcnn 网络中抠出来的 backbone,网络带 fpn,供与参考。
网络交叉
实验目的
- 探究 教师与学生采用不同网络结构 对 STPM 模型性能的影响
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
Resnet50 - WideResnet50, WideResnet50 - Resnet50, Resnet18 - Resnet34, Resnet34 - Resnet18 |
尺寸 |
256×256 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Resnet18-Resnet34 |
0.858 |
0.859 |
0.861 |
0.861 |
0.839 |
0.871 |
0.867 |
0.868 |
0.867 |
0.861 |
0.865 |
0.862 |
0.864 |
0.862 |
0.863 |
0.900 |
0.901 |
0.900 |
0.901 |
0.897 |
0.892 |
0.896 |
0.889 |
0.894 |
0.883 |
0.897 |
0.901 |
0.901 |
0.905 |
0.893 |
Resnet18-Resnet18 |
0.853 |
0.855 |
0.855 |
0.859 |
0.828 |
0.887 |
0.885 |
0.887 |
0.885 |
0.870 |
0.883 |
0.886 |
0.883 |
0.886 |
0.869 |
0.906 |
0.910 |
0.909 |
0.910 |
0.904 |
0.905 |
0.909 |
0.898 |
0.906 |
0.886 |
0.909 |
0.915 |
0.914 |
0.918 |
0.898 |
Resnet34-Resnet18 |
0.826 |
0.832 |
0.830 |
0.833 |
0.814 |
0.847 |
0.841 |
0.842 |
0.840 |
0.838 |
0.860 |
0.862 |
0.864 |
0.863 |
0.857 |
0.892 |
0.898 |
0.896 |
0.898 |
0.886 |
0.882 |
0.885 |
0.878 |
0.882 |
0.872 |
0.887 |
0.893 |
0.892 |
0.896 |
0.881 |
Resnet34-Resnet34 |
0.857 |
0.861 |
0.860 |
0.862 |
0.829 |
0.883 |
0.884 |
0.882 |
0.885 |
0.860 |
0.890 |
0.892 |
0.891 |
0.893 |
0.878 |
0.903 |
0.909 |
0.905 |
0.910 |
0.897 |
0.906 |
0.911 |
0.896 |
0.905 |
0.883 |
0.906 |
0.914 |
0.910 |
0.917 |
0.893 |
Resnet50-Resnet50Wide |
0.862 |
0.861 |
0.861 |
0.859 |
0.822 |
0.880 |
0.880 |
0.878 |
0.878 |
0.855 |
0.883 |
0.877 |
0.884 |
0.878 |
0.875 |
0.881 |
0.886 |
0.885 |
0.886 |
0.869 |
0.892 |
0.893 |
0.885 |
0.890 |
0.870 |
0.890 |
0.889 |
0.892 |
0.890 |
0.875 |
Resnet50-Resnet50 |
0.865 |
0.866 |
0.867 |
0.866 |
0.826 |
0.893 |
0.897 |
0.894 |
0.896 |
0.857 |
0.889 |
0.892 |
0.889 |
0.893 |
0.867 |
0.890 |
0.895 |
0.895 |
0.897 |
0.879 |
0.903 |
0.906 |
0.889 |
0.901 |
0.870 |
0.898 |
0.902 |
0.898 |
0.902 |
0.871 |
Resnet50Wide-Resnet50 |
0.859 |
0.867 |
0.865 |
0.868 |
0.824 |
0.863 |
0.858 |
0.859 |
0.856 |
0.850 |
0.856 |
0.859 |
0.859 |
0.860 |
0.852 |
0.884 |
0.883 |
0.886 |
0.882 |
0.869 |
0.880 |
0.883 |
0.874 |
0.881 |
0.859 |
0.879 |
0.881 |
0.881 |
0.882 |
0.863 |
WideResnet50-WideResnet50 |
0.870 |
0.881 |
0.878 |
0.884 |
0.833 |
0.888 |
0.890 |
0.889 |
0.892 |
0.852 |
0.889 |
0.892 |
0.887 |
0.890 |
0.865 |
0.908 |
0.909 |
0.910 |
0.907 |
0.897 |
0.906 |
0.911 |
0.888 |
0.903 |
0.869 |
0.907 |
0.911 |
0.906 |
0.909 |
0.875 |

相关结论
数据输入策略
实验目的
- 探究 **
教师-学生
网络分别输入原图-底板
/ 底板- 原图
**对 STPM 模型性能的影响
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
输入策略 |
原图-底板 / 底板- 原图 |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Tea-image-Stu-temp |
0.695 |
0.698 |
0.700 |
0.708 |
0.692 |
0.766 |
0.771 |
0.779 |
0.794 |
0.754 |
0.745 |
0.749 |
0.753 |
0.764 |
0.735 |
0.806 |
0.806 |
0.806 |
0.808 |
0.808 |
0.780 |
0.784 |
0.783 |
0.789 |
0.771 |
0.797 |
0.800 |
0.804 |
0.813 |
0.787 |
Tea-temp-Stu-image |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制条形图

相关结论
- 教师网络输入底板,学生网络输入原图的策略性能更优。
损失函数
实验目的
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss, cos similarity, Huber Loss, L1 |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
L2 Loss |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
cos similarity |
0.871 |
0.887 |
0.886 |
0.901 |
0.834 |
0.910 |
0.926 |
0.926 |
0.932 |
0.863 |
0.906 |
0.913 |
0.908 |
0.914 |
0.862 |
0.897 |
0.901 |
0.902 |
0.905 |
0.871 |
0.917 |
0.927 |
0.897 |
0.920 |
0.869 |
0.914 |
0.922 |
0.917 |
0.925 |
0.866 |
Huber Loss |
0.869 |
0.885 |
0.885 |
0.900 |
0.833 |
0.909 |
0.925 |
0.925 |
0.932 |
0.861 |
0.904 |
0.912 |
0.907 |
0.914 |
0.859 |
0.895 |
0.898 |
0.900 |
0.902 |
0.869 |
0.916 |
0.926 |
0.895 |
0.919 |
0.866 |
0.913 |
0.921 |
0.916 |
0.924 |
0.863 |
L1 Loss |
0.883 |
0.904 |
0.903 |
0.919 |
0.844 |
0.910 |
0.914 |
0.918 |
0.920 |
0.883 |
0.899 |
0.906 |
0.903 |
0.911 |
0.878 |
0.859 |
0.849 |
0.873 |
0.855 |
0.878 |
0.918 |
0.919 |
0.908 |
0.918 |
0.888 |
0.912 |
0.916 |
0.913 |
0.921 |
0.882 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制折线图

相关结论
- L2, Huber, Cos Similarity 损失函数得到的模型性能接近
- L1 损失会 提升第一层特征 的性能,高层特征性能下降
定位精度
实验目的
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
定位精度 |
1-2 pixel, 4-8 pixel |
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Fine |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
Coarse |
0.720 |
0.720 |
0.720 |
0.721 |
0.720 |
0.723 |
0.723 |
0.723 |
0.723 |
0.723 |
0.722 |
0.722 |
0.722 |
0.722 |
0.724 |
0.720 |
0.720 |
0.722 |
0.721 |
0.726 |
0.721 |
0.721 |
0.722 |
0.721 |
0.724 |
0.721 |
0.722 |
0.722 |
0.722 |
0.725 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制柱状图

相关结论
- STPM 通过底板进行异常检测的过程依赖较为精确的定位,如果所有数据匹配均存在超过 4 像素偏差则会严重影响算法性能
剪枝训练
实验目的
- 探究 仅训练前若干层网络 对 STPM 模型性能的影响
- 仅训练浅层网络是否会影响该部分浅层网络性能
- 以此作为网络剪枝训练、部署的依据
实验设置
实验项目 |
实验设置 |
数据集 |
Total |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
训练层级 |
layer1,layer2,layer3,layer4(BaseLine) |
- 训练层级按照网络 Stage 定义,以 Resnet 50 为例

- 定义层级后,仅该层级以及之前层级的参数可训练,其余参数冻结。
实验结果
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Level 1 |
0.873 |
0.892 |
0.889 |
0.904 |
0.834 |
0.806 |
0.810 |
0.822 |
0.827 |
0.747 |
0.646 |
0.646 |
0.656 |
0.670 |
0.651 |
0.717 |
0.730 |
0.749 |
0.753 |
0.758 |
0.875 |
0.889 |
0.887 |
0.902 |
0.826 |
0.751 |
0.738 |
0.752 |
0.768 |
0.740 |
Level 2 |
0.871 |
0.889 |
0.888 |
0.903 |
0.833 |
0.907 |
0.921 |
0.921 |
0.927 |
0.857 |
0.646 |
0.645 |
0.655 |
0.676 |
0.646 |
0.719 |
0.720 |
0.749 |
0.752 |
0.731 |
0.900 |
0.917 |
0.904 |
0.920 |
0.854 |
0.827 |
0.828 |
0.848 |
0.860 |
0.805 |
Level 3 |
0.874 |
0.890 |
0.890 |
0.905 |
0.835 |
0.912 |
0.926 |
0.926 |
0.933 |
0.864 |
0.907 |
0.914 |
0.910 |
0.917 |
0.863 |
0.724 |
0.712 |
0.761 |
0.767 |
0.748 |
0.915 |
0.926 |
0.902 |
0.925 |
0.864 |
0.903 |
0.911 |
0.916 |
0.923 |
0.857 |
Level 4 |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制条形图

相关结论
- 深层训练在浅层网络上的表现与浅层训练的网络一致
- 可以在深层网络上训练,采用浅层网络结果用于模型部署
- 可以仅在浅层网络上训练,完成后直接部署
-
数据微调
实验由来
- STPM 模型基础教师网络为与项目数据不相关的 ImageNet 等数据集预训练而来,如果教师网络适应项目数据是否会提升模型性能?
实验目的
- 探究 使用项目数据微调教师模型 对 STPM 模型数据迁移能力的影响
实验设置
实验项目 |
实验设置 |
数据集 |
TI |
模型 |
Resnet50 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
SSL |
损失函数 |
L2 Loss |
教师网络预训练 |
TI 数据有监督分类任务 |
冻结参数 |
Layer1 |
- 使用人工标注的二分类数据预训练教师网络,二分类 AP 由 0.6+ 提升到了0.9,之后保存模型用于STPM模型训练
- 随后实验发现微调过的模型浅层特征提取能力会下降,因此冻结了浅层特征又做了一组对比实验
实验结果
- 采用在
Part1, Part2, Part3, Total
数据 测试集 上的 二分类 AP 作为评判指标:
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
baseline |
0.933 |
0.950 |
0.954 |
0.958 |
0.889 |
0.853 |
0.890 |
0.897 |
0.929 |
0.814 |
0.730 |
0.758 |
0.763 |
0.805 |
0.693 |
0.730 |
0.745 |
0.707 |
0.786 |
0.596 |
0.849 |
0.877 |
0.819 |
0.888 |
0.750 |
0.745 |
0.765 |
0.730 |
0.822 |
0.627 |
finetune |
0.873 |
0.897 |
0.899 |
0.910 |
0.845 |
0.881 |
0.899 |
0.899 |
0.912 |
0.871 |
0.811 |
0.832 |
0.833 |
0.847 |
0.799 |
0.750 |
0.777 |
0.763 |
0.799 |
0.652 |
0.869 |
0.886 |
0.871 |
0.889 |
0.841 |
0.770 |
0.791 |
0.782 |
0.819 |
0.730 |
finetune freeze_L1 |
0.940 |
0.950 |
0.953 |
0.956 |
0.842 |
0.855 |
0.890 |
0.903 |
0.931 |
0.816 |
0.772 |
0.789 |
0.796 |
0.825 |
0.759 |
0.738 |
0.766 |
0.764 |
0.805 |
0.712 |
0.875 |
0.891 |
0.872 |
0.905 |
0.809 |
0.753 |
0.776 |
0.782 |
0.828 |
0.734 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)绘制条形图

相关结论
- 直接使用项目数据微调模型,可以显著提升深层网络(
layer3, 4
)特征性能,但会损失浅层特征的性能(layer1
)
- 为了解决浅层特征性能下降的问题冻结浅层(
layer1
)后深层特征相比 baseline 均有性能提升,但 layer3
层性能不如直接微调的收益高
特征组合策略
实验由来
- 在
数据3
中发现在跨底板迁移时使用直接四层特征相加时阈值不稳定,而且精度有限,尝试使用其他的阈值判定策略。
实验目的
- 探究 特征组合策略 对 STPM 模型数据迁移能力的影响
实验设置
实验项目 |
实验设置 |
数据集 |
数据3 —— 9套TI底板数据 |
模型 |
Wide-Resnet50 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
实验结果
- 采用在
数据3
测试集 上的 二分类 AP 作为评判指标:
NG AP |
center_weight_sum_1_3.json |
gauss_mul_1_3.json |
gauss_sum_1_3.json |
gauss_weight_sum_1_3.json |
L1_Gau_L |
L2_Gau_L |
L3_Gau_L |
L4_Gau_L |
lev_mul_1_3.json |
lev_sum_1_3.json |
Mul_CC_L |
Mul_G_L |
Mul_N |
sigma_num_mul.json |
sigma_num_sum.json |
Sum_CC_L |
Sum_G_L |
Sum_N |
weight_sum_1_3 |
MH-BALLDROP |
0.851 |
0.891 |
0.912 |
0.919 |
0.766 |
0.903 |
0.731 |
0.318 |
0.774 |
0.824 |
0.641 |
0.680 |
0.360 |
0.469 |
0.727 |
0.545 |
0.660 |
0.141 |
0.827 |
MH-M1 |
0.638 |
0.709 |
0.747 |
0.735 |
0.692 |
0.730 |
0.526 |
0.308 |
0.597 |
0.626 |
0.556 |
0.546 |
0.298 |
0.672 |
0.722 |
0.448 |
0.505 |
0.207 |
0.612 |
MH-UBM |
0.872 |
0.821 |
0.886 |
0.859 |
0.930 |
0.880 |
0.732 |
0.771 |
0.843 |
0.878 |
0.819 |
0.810 |
0.751 |
0.553 |
0.772 |
0.790 |
0.761 |
0.678 |
0.872 |
RF-BALLDROP |
0.408 |
0.420 |
0.462 |
0.454 |
0.215 |
0.469 |
0.406 |
0.080 |
0.284 |
0.370 |
0.304 |
0.300 |
0.071 |
0.442 |
0.556 |
0.225 |
0.318 |
0.041 |
0.366 |
RF-M1 |
0.825 |
0.716 |
0.890 |
0.904 |
0.228 |
0.894 |
0.746 |
0.416 |
0.578 |
0.773 |
0.587 |
0.608 |
0.268 |
0.094 |
0.788 |
0.596 |
0.728 |
0.027 |
0.765 |
RF-UBM |
0.832 |
0.895 |
0.907 |
0.898 |
0.815 |
0.887 |
0.801 |
0.266 |
0.846 |
0.842 |
0.774 |
0.776 |
0.408 |
0.648 |
0.882 |
0.661 |
0.727 |
0.134 |
0.820 |
BALLDROP |
0.949 |
0.973 |
0.964 |
0.969 |
0.897 |
0.963 |
0.980 |
0.901 |
0.918 |
0.905 |
0.977 |
0.960 |
0.818 |
0.615 |
0.755 |
0.973 |
0.981 |
0.553 |
0.922 |
M1 |
0.871 |
0.961 |
0.972 |
0.969 |
0.908 |
0.969 |
0.856 |
0.408 |
0.796 |
0.780 |
0.825 |
0.765 |
0.244 |
0.605 |
0.894 |
0.741 |
0.829 |
0.099 |
0.800 |
UBM |
0.911 |
0.929 |
0.915 |
0.908 |
0.914 |
0.889 |
0.819 |
0.833 |
0.808 |
0.908 |
0.934 |
0.894 |
0.729 |
0.322 |
0.502 |
0.893 |
0.893 |
0.578 |
0.889 |

相关结论
- 经过多种特征组合实验,发现第一层特征辅助,第二层特征为主,结合第三层特征,放弃第四层特征的策略比较有效,无论在 AP 上还是召回率上
Padding 训练
实验目的
- 探究训练模型时对原始数据添加 Padding 预处理操作对 STPM 性能的影像
实验设置
实验项目 |
实验设置 |
数据集 |
数据3 —— 9套TI底板数据 |
模型 |
Wide-Resnet50, Resnet50, Resnet18 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
实验结果
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 作为评判指标:
Wide-Resnet 50
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean |
no padding |
0.919 |
0.735 |
0.859 |
0.454 |
0.904 |
0.898 |
0.969 |
0.969 |
0.908 |
0.846 |
ref padding |
0.922 |
0.767 |
0.875 |
0.480 |
0.920 |
0.919 |
0.971 |
0.970 |
0.928 |
0.861 |
const 0 padding |
0.920 |
0.763 |
0.874 |
0.455 |
0.909 |
0.908 |
0.970 |
0.974 |
0.913 |
0.854 |
const 128 padding |
0.912 |
0.726 |
0.859 |
0.466 |
0.900 |
0.905 |
0.969 |
0.971 |
0.921 |
0.848 |
const 255 padding |
0.913 |
0.766 |
0.870 |
0.476 |
0.929 |
0.909 |
0.970 |
0.974 |
0.916 |
0.858 |
Resnet 50
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean |
no padding |
0.911 |
0.758 |
0.907 |
0.398 |
0.969 |
0.937 |
0.952 |
0.975 |
0.950 |
0.862 |
ref padding |
0.919 |
0.784 |
0.918 |
0.414 |
0.969 |
0.941 |
0.958 |
0.979 |
0.956 |
0.871 |
const 0 padding |
0.914 |
0.754 |
0.921 |
0.387 |
0.976 |
0.939 |
0.957 |
0.978 |
0.943 |
0.863 |
const 128 padding |
0.914 |
0.768 |
0.914 |
0.392 |
0.979 |
0.941 |
0.953 |
0.975 |
0.961 |
0.866 |
const 255 padding |
0.918 |
0.771 |
0.920 |
0.398 |
0.975 |
0.942 |
0.958 |
0.975 |
0.960 |
0.869 |
Resnet 18
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean |
no padding |
0.850 |
0.679 |
0.899 |
0.368 |
0.931 |
0.898 |
0.922 |
0.938 |
0.910 |
0.822 |
ref padding |
0.865 |
0.718 |
0.911 |
0.377 |
0.949 |
0.905 |
0.932 |
0.942 |
0.921 |
0.836 |
const 0 padding |
0.864 |
0.677 |
0.915 |
0.366 |
0.967 |
0.902 |
0.926 |
0.938 |
0.875 |
0.826 |
const 128 padding |
0.857 |
0.690 |
0.912 |
0.354 |
0.952 |
0.893 |
0.922 |
0.939 |
0.910 |
0.826 |
const 255 padding |
0.865 |
0.697 |
0.915 |
0.353 |
0.951 |
0.896 |
0.924 |
0.939 |
0.916 |
0.828 |
结果汇总
- 将三种网络结构在不同 Padding 策略下的九种数据集 NG AP 的均值绘制成折线图:

相关结论
- 在训练阶段加入 Padding 操作可以提升 STPM 网络性能
- 反射 Padding 相比于 Constant Padding 提示性能幅度更加明显
- Constant Padding 中,相比于 0 和 128, 值采用 255 更能提升模型性能
- 当前一个经验结论是反射 Padding 可以一定程度上抑制 STPM 模型在黑暗区域的异常响应
模型迁移能力 测试一
实验目的
实验设置
实验项目 |
实验设置 |
数据集 |
Total, Part1, Part2, Part3 |
模型 |
WideResnet50 |
尺寸 |
600×600 |
Batch Size |
6 |
Epoch |
36 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
训练层级 |
layer4 |
- 分别用四种数据集作训练集、测试集训练模型,评估结果
实验结果
- 采用在不同大类别 测试集 上的 二分类 AP 作为评判指标
NG AP |
L1_CC_L |
L1_CC_S |
L1_Gau_L |
L1_Gau_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_Gau_L |
L2_Gau_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_Gau_L |
L3_Gau_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_Gau_L |
L4_Gau_S |
L4_N |
Mul_CC_L |
Mul_CC_S |
Mul_Gau_L |
Mul_Gau_S |
Mul_N |
Sum_CC_L |
Sum_CC_S |
Sum_Gau_L |
Sum_Gau_S |
Sum_N |
Train Part1-Test Total |
0.748 |
0.747 |
0.751 |
0.762 |
0.742 |
0.811 |
0.828 |
0.833 |
0.844 |
0.784 |
0.752 |
0.759 |
0.760 |
0.765 |
0.739 |
0.712 |
0.731 |
0.731 |
0.733 |
0.714 |
0.750 |
0.757 |
0.754 |
0.759 |
0.738 |
0.738 |
0.753 |
0.753 |
0.759 |
0.719 |
Train Part2-Test Total |
0.780 |
0.799 |
0.805 |
0.821 |
0.765 |
0.775 |
0.803 |
0.808 |
0.828 |
0.754 |
0.727 |
0.741 |
0.743 |
0.752 |
0.716 |
0.679 |
0.675 |
0.681 |
0.682 |
0.675 |
0.722 |
0.734 |
0.731 |
0.739 |
0.709 |
0.702 |
0.708 |
0.713 |
0.719 |
0.689 |
Train Part3-Test Total |
0.853 |
0.866 |
0.873 |
0.880 |
0.826 |
0.882 |
0.890 |
0.892 |
0.897 |
0.863 |
0.864 |
0.863 |
0.867 |
0.869 |
0.856 |
0.816 |
0.817 |
0.815 |
0.818 |
0.814 |
0.857 |
0.863 |
0.863 |
0.868 |
0.850 |
0.836 |
0.845 |
0.848 |
0.855 |
0.836 |
Train Total-Test Total |
0.870 |
0.888 |
0.888 |
0.902 |
0.833 |
0.910 |
0.926 |
0.926 |
0.932 |
0.862 |
0.906 |
0.913 |
0.908 |
0.915 |
0.861 |
0.896 |
0.900 |
0.900 |
0.903 |
0.870 |
0.917 |
0.927 |
0.896 |
0.921 |
0.868 |
0.915 |
0.922 |
0.917 |
0.925 |
0.865 |
Train Part1-Test Part2 |
0.805 |
0.795 |
0.801 |
0.813 |
0.809 |
0.894 |
0.909 |
0.912 |
0.917 |
0.882 |
0.835 |
0.849 |
0.844 |
0.858 |
0.843 |
0.797 |
0.847 |
0.839 |
0.862 |
0.824 |
0.858 |
0.876 |
0.869 |
0.877 |
0.863 |
0.832 |
0.861 |
0.854 |
0.876 |
0.843 |
Train Part2-Test Part2 |
0.864 |
0.867 |
0.867 |
0.886 |
0.871 |
0.898 |
0.924 |
0.924 |
0.935 |
0.890 |
0.889 |
0.905 |
0.901 |
0.909 |
0.887 |
0.900 |
0.904 |
0.911 |
0.908 |
0.908 |
0.906 |
0.920 |
0.917 |
0.918 |
0.905 |
0.901 |
0.914 |
0.913 |
0.923 |
0.901 |
Train Part3-Test Part2 |
0.822 |
0.833 |
0.842 |
0.850 |
0.824 |
0.834 |
0.851 |
0.855 |
0.863 |
0.823 |
0.812 |
0.814 |
0.817 |
0.820 |
0.818 |
0.779 |
0.787 |
0.781 |
0.783 |
0.780 |
0.809 |
0.822 |
0.821 |
0.827 |
0.814 |
0.785 |
0.803 |
0.803 |
0.811 |
0.800 |
Train Total-Test Part2 |
0.881 |
0.879 |
0.882 |
0.900 |
0.887 |
0.907 |
0.931 |
0.931 |
0.939 |
0.896 |
0.891 |
0.906 |
0.901 |
0.909 |
0.888 |
0.902 |
0.904 |
0.912 |
0.909 |
0.900 |
0.906 |
0.921 |
0.917 |
0.920 |
0.906 |
0.901 |
0.915 |
0.914 |
0.923 |
0.900 |
Train Part1-Test Part1 |
0.916 |
0.937 |
0.937 |
0.949 |
0.902 |
0.960 |
0.969 |
0.969 |
0.976 |
0.936 |
0.959 |
0.965 |
0.963 |
0.967 |
0.936 |
0.962 |
0.965 |
0.962 |
0.967 |
0.946 |
0.966 |
0.970 |
0.952 |
0.968 |
0.938 |
0.965 |
0.970 |
0.969 |
0.973 |
0.941 |
Train Part2-Test Part1 |
0.918 |
0.939 |
0.939 |
0.949 |
0.890 |
0.950 |
0.962 |
0.961 |
0.968 |
0.920 |
0.948 |
0.947 |
0.951 |
0.948 |
0.919 |
0.904 |
0.909 |
0.909 |
0.923 |
0.905 |
0.945 |
0.951 |
0.943 |
0.952 |
0.914 |
0.932 |
0.935 |
0.938 |
0.945 |
0.921 |
Train Part3-Test Part1 |
0.902 |
0.917 |
0.921 |
0.925 |
0.897 |
0.937 |
0.936 |
0.939 |
0.942 |
0.934 |
0.932 |
0.927 |
0.932 |
0.933 |
0.929 |
0.876 |
0.876 |
0.879 |
0.883 |
0.880 |
0.924 |
0.925 |
0.928 |
0.931 |
0.921 |
0.911 |
0.911 |
0.916 |
0.921 |
0.910 |
Train Total-Test Part1 |
0.923 |
0.938 |
0.937 |
0.945 |
0.901 |
0.959 |
0.966 |
0.966 |
0.971 |
0.935 |
0.957 |
0.960 |
0.960 |
0.962 |
0.934 |
0.956 |
0.957 |
0.957 |
0.959 |
0.943 |
0.961 |
0.965 |
0.951 |
0.963 |
0.936 |
0.962 |
0.965 |
0.965 |
0.966 |
0.940 |
Train Part1-Test Part3 |
0.471 |
0.477 |
0.484 |
0.504 |
0.421 |
0.646 |
0.655 |
0.662 |
0.675 |
0.574 |
0.620 |
0.626 |
0.636 |
0.650 |
0.546 |
0.525 |
0.548 |
0.553 |
0.541 |
0.459 |
0.633 |
0.631 |
0.624 |
0.639 |
0.534 |
0.611 |
0.614 |
0.624 |
0.626 |
0.506 |
Train Part2-Test Part3 |
0.578 |
0.620 |
0.634 |
0.661 |
0.449 |
0.672 |
0.686 |
0.693 |
0.686 |
0.617 |
0.579 |
0.597 |
0.604 |
0.615 |
0.532 |
0.400 |
0.406 |
0.415 |
0.432 |
0.383 |
0.605 |
0.624 |
0.614 |
0.635 |
0.509 |
0.531 |
0.547 |
0.558 |
0.574 |
0.453 |
Train Part3-Test Part3 |
0.685 |
0.745 |
0.743 |
0.772 |
0.528 |
0.695 |
0.714 |
0.707 |
0.719 |
0.612 |
0.723 |
0.737 |
0.717 |
0.736 |
0.606 |
0.774 |
0.780 |
0.759 |
0.781 |
0.573 |
0.762 |
0.773 |
0.685 |
0.749 |
0.602 |
0.769 |
0.769 |
0.752 |
0.766 |
0.593 |
Train Total-Test Part3 |
0.681 |
0.739 |
0.734 |
0.763 |
0.524 |
0.699 |
0.719 |
0.711 |
0.718 |
0.614 |
0.729 |
0.740 |
0.722 |
0.736 |
0.606 |
0.772 |
0.779 |
0.764 |
0.779 |
0.570 |
0.764 |
0.775 |
0.693 |
0.751 |
0.606 |
0.774 |
0.774 |
0.757 |
0.766 |
0.594 |
相关结论
Part1-Part3
相互作为训练、测试集的结果可以看出:
- 3×3的格子内对角线元素分值最高,表明使用和训练集不同源的数据进行测试确实会降低模型性能
- 其中
Part1, Part2
的交叉性能偏差不大,Part3
中的数据集 MissMatch 对结果的影响更大一些,而事实上 Part1, Part2
的底板形成模式确实更加接近(1PxM
),Part3
为不同的 2PxM
,表明跨底板时越是接近的底板越容易迁移
- 根据
Total
作为训练集时的结果可以看出:
- 在多组数据中 Total 训练的模型都有着最佳的表现,表明如果有多种底板来源的数据,可以训练大一统模型用于异常检测,而不会带来性能损失,甚至会有些收益
- 根据四层特征结果表格的结果可以看出:
Layer 1
和 Layer 4
的特征受数据 MissMatch 的影响更大
Layer 2
和 Layer 3
的特征结果在 MissMatch 数据上的表现较好(尤其是 Layer 2
),没有严重的降点
- 跨底板迁移时应以
Layer 2,3
为主要应用的特征层
- 四层特征在变化趋势上基本一致
模型迁移能力 测试二
实验目的
- 在
数据3
上探究 STPM 模型数据迁移能力
- 分别用每种数据训练,所有(9种)数据测试得到实验结果
实验设置
实验项目 |
实验设置 |
数据集 |
数据3 —— 9套TI底板数据 |
模型 |
Wide-Resnet50 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
实验结果
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 作为评判指标:
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BallDrop |
0.918 |
0.485 |
0.829 |
0.458 |
0.765 |
0.876 |
0.936 |
0.824 |
0.938 |
MH-M1 |
0.880 |
0.656 |
0.661 |
0.357 |
0.889 |
0.850 |
0.903 |
0.717 |
0.412 |
MH-UBM |
0.630 |
0.204 |
0.874 |
0.123 |
0.477 |
0.838 |
0.746 |
0.687 |
0.713 |
RF-BALLDROP |
0.904 |
0.672 |
0.845 |
0.489 |
0.728 |
0.884 |
0.875 |
0.953 |
0.956 |
RF-M1 |
0.857 |
0.607 |
0.773 |
0.333 |
0.941 |
0.902 |
0.849 |
0.888 |
0.779 |
RF-UBM |
0.813 |
0.515 |
0.787 |
0.270 |
0.848 |
0.892 |
0.905 |
0.896 |
0.566 |
BALLDROP |
0.865 |
0.599 |
0.918 |
0.361 |
0.926 |
0.900 |
0.969 |
0.902 |
0.707 |
M1 |
0.853 |
0.595 |
0.720 |
0.338 |
0.913 |
0.891 |
0.817 |
0.973 |
0.877 |
UBM |
0.928 |
0.464 |
0.901 |
0.477 |
0.960 |
0.934 |
0.910 |
0.954 |
0.972 |
All-Data |
0.919 |
0.735 |
0.859 |
0.454 |
0.904 |
0.898 |
0.969 |
0.969 |
0.908 |
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean Value |
MH-BallDrop |
1.000 |
0.739 |
0.949 |
0.937 |
0.813 |
0.982 |
0.966 |
0.847 |
0.965 |
0.911 |
MH-M1 |
0.959 |
1.000 |
0.756 |
0.730 |
0.945 |
0.953 |
0.932 |
0.737 |
0.424 |
0.826 |
MH-UBM |
0.686 |
0.311 |
1.000 |
0.252 |
0.507 |
0.939 |
0.770 |
0.706 |
0.734 |
0.656 |
RF-BALLDROP |
0.985 |
1.024 |
0.967 |
1.000 |
0.774 |
0.991 |
0.903 |
0.979 |
0.984 |
0.956 |
RF-M1 |
0.934 |
0.925 |
0.884 |
0.681 |
1.000 |
1.011 |
0.876 |
0.913 |
0.801 |
0.892 |
RF-UBM |
0.886 |
0.785 |
0.900 |
0.552 |
0.901 |
1.000 |
0.934 |
0.921 |
0.582 |
0.829 |
BALLDROP |
0.942 |
0.913 |
1.050 |
0.738 |
0.984 |
1.009 |
1.000 |
0.927 |
0.727 |
0.921 |
M1 |
0.929 |
0.907 |
0.824 |
0.691 |
0.970 |
0.999 |
0.843 |
1.000 |
0.902 |
0.896 |
UBM |
1.011 |
0.707 |
1.031 |
0.975 |
1.020 |
1.047 |
0.939 |
0.980 |
1.000 |
0.968 |
All-Data |
1.001 |
1.120 |
0.983 |
0.928 |
0.961 |
1.007 |
1.000 |
0.996 |
0.934 |
0.992 |
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的AP拐点NG得分作为阈值指标:
→Test ↓ Train |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
2.010614 |
2.144005 |
1.810394 |
1.881123 |
2.287157 |
1.99912 |
1.946828 |
2.185724 |
1.673338 |
MH-M1 |
2.162106 |
1.411933 |
2.029175 |
2.097367 |
2.099392 |
2.046499 |
1.971851 |
2.231965 |
2.051707 |
MH-UBM |
2.370459 |
2.009487 |
1.848498 |
2.45808 |
2.605074 |
2.402819 |
2.497695 |
2.587343 |
2.171575 |
RF-BALLDROP |
2.040084 |
2.002772 |
1.863159 |
1.726423 |
2.231972 |
1.849009 |
1.852682 |
2.098804 |
1.648024 |
RF-M1 |
2.156496 |
2.002666 |
2.025805 |
2.104126 |
1.974905 |
2.162568 |
2.19801 |
2.198959 |
1.833202 |
RF-UBM |
2.101938 |
1.988546 |
1.768906 |
2.226227 |
2.068382 |
1.750765 |
1.977585 |
2.075967 |
1.974408 |
BALLDROP |
2.192295 |
2.182644 |
2.003255 |
2.271597 |
2.172575 |
2.19438 |
1.775792 |
2.312731 |
2.139986 |
M1 |
2.101975 |
2.054031 |
2.060857 |
2.124095 |
2.059971 |
2.198254 |
2.293466 |
1.881917 |
1.875218 |
UBM |
2.17613 |
2.395239 |
1.93896 |
2.217171 |
2.373836 |
2.325385 |
2.155779 |
2.258681 |
1.422584 |
相关结论
- 视觉相似的底板事实上是可以迁移测试使用的,但相差较大的底板迁移性能下降较为严重,即 STPM 具有一定的迁移能力,但模型没有完全的迁移测试能力
- 在不同数据集上训练的模型迁移能力相差较大,但以 UBM / RF-BALLDROP 数据训练的模型迁移能力很强,归一化 AP 均值超过了 0.95
- 使用所有数据训练的模型为表中 All-Data 行,归一化 AP 均值为 0.992,基本与单数据集训练的模型性能相当,因此使用多种数据混合训练STPM模型可以作为一种贪心策略
- NG score 在所使用的特征选取策略上表现较为稳定
模型迁移能力 测试三
实验目的
- 在
数据3
上采用交叉验证的方式探究 STPM 模型数据迁移能力与训练数据量、数据集的关系
实验设置
实验项目 |
实验设置 |
数据集 |
数据3 —— 9套TI底板数据,分为三组 |
数据分组 |
A - [RF-Balldrop, RF-M1, Balldrop], B - [UBM, MH-UBM, M1], C - [MH-M1, RF-UBM, MH-Balldrop] |
模型 |
Wide-Resnet50 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
训练数据量 |
1000, 2000, 4000 (多则裁剪,少则重复) |
- 三组数据分别轮流作为 Train1,Train2,test 数据,共三组实验
实验结果
A,B 训练集,C 测试集
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 作为评判指标:
ng ap |
A-RF-BALLDROP |
A-RF-M1 |
A-BALLDROP |
A-Mean |
B-MH-UBM |
B-M1 |
B-UBM |
B-Mean |
C-MH-BALLDROP |
C-MH-M1 |
C-RF-UBM |
C-Mean |
A 1000 B 1000 |
0.476 |
0.933 |
0.980 |
0.796 |
0.913 |
0.978 |
0.940 |
0.944 |
0.910 |
0.688 |
0.914 |
0.837 |
A 2000 B 2000 |
0.498 |
0.909 |
0.984 |
0.797 |
0.915 |
0.979 |
0.954 |
0.949 |
0.919 |
0.649 |
0.910 |
0.826 |
A 4000 B 4000 |
0.500 |
0.935 |
0.984 |
0.806 |
0.915 |
0.981 |
0.964 |
0.953 |
0.903 |
0.595 |
0.891 |
0.797 |
A 1000 B 2000 |
0.470 |
0.930 |
0.980 |
0.793 |
0.926 |
0.980 |
0.959 |
0.955 |
0.911 |
0.686 |
0.910 |
0.836 |
A 1000 B 4000 |
0.479 |
0.912 |
0.980 |
0.790 |
0.920 |
0.980 |
0.965 |
0.955 |
0.917 |
0.622 |
0.917 |
0.819 |
A 2000 B 1000 |
0.497 |
0.929 |
0.982 |
0.803 |
0.903 |
0.983 |
0.944 |
0.943 |
0.916 |
0.663 |
0.906 |
0.828 |
A 4000 B 1000 |
0.489 |
0.938 |
0.981 |
0.803 |
0.903 |
0.984 |
0.938 |
0.942 |
0.914 |
0.627 |
0.900 |
0.814 |
Mean |
0.487 |
0.927 |
0.982 |
0.798 |
0.914 |
0.981 |
0.952 |
0.949 |
0.913 |
0.647 |
0.907 |
0.822 |
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 的拐点作为阈值参考值:
score |
A-RF-BALLDROP |
A-RF-M1 |
A-BALLDROP |
A-Mean |
B-MH-UBM |
B-M1 |
B-UBM |
B-Mean |
C-MH-BALLDROP |
C-MH-M1 |
C-RF-UBM |
C-Mean |
A 1000 B 1000 |
1.885 |
1.879 |
1.603 |
1.789 |
1.666 |
1.948 |
1.498 |
1.704 |
1.968 |
1.860 |
1.964 |
1.931 |
A 2000 B 2000 |
1.837 |
1.917 |
1.605 |
1.786 |
1.546 |
2.024 |
1.367 |
1.645 |
1.960 |
1.930 |
1.893 |
1.928 |
A 4000 B 4000 |
1.886 |
1.805 |
1.589 |
1.760 |
1.232 |
1.885 |
1.257 |
1.458 |
1.972 |
1.924 |
2.040 |
1.979 |
A 1000 B 2000 |
1.727 |
1.917 |
1.622 |
1.756 |
1.475 |
1.817 |
1.379 |
1.557 |
1.953 |
1.861 |
1.862 |
1.892 |
A 1000 B 4000 |
1.903 |
1.878 |
1.656 |
1.812 |
1.315 |
1.945 |
1.265 |
1.508 |
2.024 |
2.007 |
1.891 |
1.974 |
A 2000 B 1000 |
1.730 |
1.897 |
1.501 |
1.709 |
1.583 |
2.005 |
1.482 |
1.690 |
1.940 |
1.901 |
1.857 |
1.899 |
A 4000 B 1000 |
1.871 |
1.875 |
1.448 |
1.731 |
1.541 |
1.916 |
1.486 |
1.648 |
1.930 |
1.925 |
1.887 |
1.914 |
Mean |
1.834 |
1.881 |
1.575 |
1.763 |
1.480 |
1.934 |
1.391 |
1.602 |
1.964 |
1.916 |
1.913 |
1.931 |
A,C 训练集,B 测试集
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 作为评判指标:
ng ap |
A-RF-BALLDROP |
A-RF-M1 |
A-BALLDROP |
A-Mean |
B-MH-UBM |
B-M1 |
B-UBM |
B-Mean |
C-MH-BALLDROP |
C-MH-M1 |
C-RF-UBM |
C-Mean |
A 1000 C 1000 |
0.456 |
0.897 |
0.965 |
0.773 |
0.774 |
0.931 |
0.831 |
0.845 |
0.917 |
0.727 |
0.889 |
0.845 |
A 2000 C 2000 |
0.467 |
0.893 |
0.968 |
0.776 |
0.747 |
0.931 |
0.786 |
0.821 |
0.917 |
0.737 |
0.879 |
0.845 |
A 4000 C 4000 |
0.499 |
0.927 |
0.979 |
0.802 |
0.726 |
0.884 |
0.712 |
0.774 |
0.907 |
0.606 |
0.869 |
0.794 |
A 1000 C 2000 |
0.458 |
0.871 |
0.965 |
0.765 |
0.734 |
0.904 |
0.768 |
0.802 |
0.917 |
0.712 |
0.880 |
0.836 |
A 1000 C 4000 |
0.462 |
0.840 |
0.973 |
0.758 |
0.678 |
0.854 |
0.597 |
0.710 |
0.920 |
0.645 |
0.858 |
0.808 |
A 2000 C 1000 |
0.460 |
0.887 |
0.971 |
0.773 |
0.784 |
0.939 |
0.799 |
0.840 |
0.910 |
0.727 |
0.886 |
0.841 |
A 4000 C 1000 |
0.486 |
0.920 |
0.976 |
0.794 |
0.777 |
0.933 |
0.873 |
0.861 |
0.907 |
0.735 |
0.892 |
0.845 |
Mean |
0.470 |
0.891 |
0.971 |
0.777 |
0.746 |
0.911 |
0.767 |
0.808 |
0.914 |
0.699 |
0.879 |
0.831 |
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 的拐点作为阈值参考值:
score |
A-RF-BALLDROP |
A-RF-M1 |
A-BALLDROP |
A-Mean |
B-MH-UBM |
B-M1 |
B-UBM |
B-Mean |
C-MH-BALLDROP |
C-MH-M1 |
C-RF-UBM |
C-Mean |
A 1000 C 1000 |
1.730 |
1.850 |
1.608 |
1.729 |
1.733 |
1.962 |
1.662 |
1.786 |
1.889 |
1.518 |
1.682 |
1.696 |
A 2000 C 2000 |
1.794 |
1.890 |
1.540 |
1.742 |
1.714 |
1.886 |
1.716 |
1.772 |
1.878 |
1.444 |
1.655 |
1.659 |
A 4000 C 4000 |
1.667 |
1.850 |
1.485 |
1.667 |
1.723 |
1.833 |
1.705 |
1.753 |
1.861 |
1.447 |
1.590 |
1.633 |
A 1000 C 2000 |
1.837 |
1.972 |
1.708 |
1.839 |
1.708 |
1.963 |
1.719 |
1.797 |
1.873 |
1.438 |
1.680 |
1.663 |
A 1000 C 4000 |
1.741 |
1.941 |
1.667 |
1.783 |
1.582 |
1.877 |
1.819 |
1.759 |
1.870 |
1.464 |
1.696 |
1.677 |
A 2000 C 1000 |
1.797 |
1.906 |
1.535 |
1.746 |
1.662 |
1.915 |
1.734 |
1.770 |
1.931 |
1.481 |
1.707 |
1.706 |
A 4000 C 1000 |
1.663 |
1.909 |
1.535 |
1.702 |
1.685 |
1.992 |
1.635 |
1.771 |
1.938 |
1.470 |
1.737 |
1.715 |
Mean |
1.747 |
1.903 |
1.583 |
1.744 |
1.687 |
1.918 |
1.713 |
1.773 |
1.891 |
1.466 |
1.678 |
1.678 |
B,C 训练集,A 测试集
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 作为评判指标:
ng ap |
A-RF-BALLDROP |
A-RF-M1 |
A-BALLDROP |
A-Mean |
B-MH-UBM |
B-M1 |
B-UBM |
B-Mean |
C-MH-BALLDROP |
C-MH-M1 |
C-RF-UBM |
C-Mean |
B 1000 C 1000 |
0.443 |
0.896 |
0.946 |
0.762 |
0.857 |
0.983 |
0.910 |
0.917 |
0.923 |
0.722 |
0.901 |
0.849 |
B 2000 C 2000 |
0.455 |
0.885 |
0.950 |
0.763 |
0.854 |
0.955 |
0.925 |
0.911 |
0.929 |
0.684 |
0.887 |
0.833 |
B 4000 C 4000 |
0.459 |
0.926 |
0.935 |
0.773 |
0.873 |
0.944 |
0.942 |
0.920 |
0.922 |
0.626 |
0.888 |
0.812 |
B 1000 C 2000 |
0.448 |
0.884 |
0.944 |
0.759 |
0.826 |
0.959 |
0.901 |
0.895 |
0.920 |
0.706 |
0.890 |
0.839 |
B 1000 C 4000 |
0.441 |
0.875 |
0.945 |
0.753 |
0.832 |
0.933 |
0.885 |
0.883 |
0.922 |
0.637 |
0.870 |
0.810 |
B 2000 C 1000 |
0.439 |
0.893 |
0.946 |
0.759 |
0.879 |
0.975 |
0.928 |
0.928 |
0.918 |
0.733 |
0.906 |
0.852 |
B 4000 C 1000 |
0.441 |
0.939 |
0.942 |
0.774 |
0.899 |
0.985 |
0.950 |
0.944 |
0.925 |
0.687 |
0.917 |
0.843 |
Mean |
0.447 |
0.900 |
0.944 |
0.763 |
0.860 |
0.962 |
0.920 |
0.914 |
0.923 |
0.685 |
0.894 |
0.834 |
- 采用
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 的拐点作为阈值参考值:
score |
A-RF-BALLDROP |
A-RF-M1 |
A-BALLDROP |
A-Mean |
B-MH-UBM |
B-M1 |
B-UBM |
B-Mean |
C-MH-BALLDROP |
C-MH-M1 |
C-RF-UBM |
C-Mean |
B 1000 C 1000 |
1.690 |
1.902 |
1.846 |
1.813 |
1.499 |
1.786 |
1.480 |
1.588 |
1.903 |
1.533 |
1.806 |
1.747 |
B 2000 C 2000 |
1.691 |
1.892 |
1.750 |
1.778 |
1.318 |
1.716 |
1.376 |
1.470 |
1.865 |
1.451 |
1.689 |
1.668 |
B 4000 C 4000 |
1.908 |
1.988 |
1.900 |
1.932 |
1.150 |
1.755 |
1.113 |
1.340 |
1.898 |
1.473 |
1.714 |
1.695 |
B 1000 C 2000 |
1.860 |
1.998 |
1.834 |
1.898 |
1.530 |
1.764 |
1.454 |
1.583 |
1.917 |
1.500 |
1.654 |
1.690 |
B 1000 C 4000 |
1.816 |
1.870 |
1.803 |
1.830 |
1.469 |
1.767 |
1.425 |
1.554 |
1.852 |
1.476 |
1.600 |
1.642 |
B 2000 C 1000 |
1.851 |
2.050 |
1.850 |
1.917 |
1.347 |
1.723 |
1.413 |
1.494 |
1.934 |
1.523 |
1.731 |
1.729 |
B 4000 C 1000 |
1.821 |
1.898 |
2.006 |
1.908 |
1.242 |
1.752 |
1.223 |
1.406 |
1.924 |
1.513 |
1.908 |
1.782 |
Mean |
1.805 |
1.942 |
1.856 |
1.868 |
1.365 |
1.752 |
1.355 |
1.491 |
1.899 |
1.496 |
1.729 |
1.708 |
综合结果
-
取三组实验结果的NG AP平均值:
行为 两种训练集训练,列为测试集
NG AP |
A |
B |
C |
AB |
0.798 |
0.949 |
0.822 |
AC |
0.777 |
0.808 |
0.831 |
BC |
0.763 |
0.914 |
0.834 |

相关结论
- STPM 训练数据量超过 1000 后增加数据量不一定会有性能收益
- 数据 A,C 迁移测试时表现相对稳定,B 测试迁移能力时 AP 下降较为严重,可能需要设计数据的距离度量来预测数据迁移能力
- AP 拐点并不稳定,和数据量没有看到明显相关性
模型迁移能力 测试四
实验目的
- 在
数据3
上探究 STPM 模型数据迁移能力
- 在
数据3
选择出制程和产品的数据作为测试集(共6种方案),数据3
其余数据作为训练集,结合 数据1
作为训练集时训练集为 Universe
,不结合 数据1
时训练集称为 All
,训练出模型在对应测试集测试得到实验结果
实验设置
实验项目 |
实验设置 |
数据集 |
数据1,数据3 —— 9套TI底板数据 |
模型 |
Wide-Resnet50 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
预训练数据 |
ImageNet |
损失函数 |
L2 Loss |
实验结果
- 采用挑选
数据3
测试集 的 Gaussian-Weighted-1-3_sum
特征提取策略下的 二分类 AP 作为评判指标:
Universe -
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
BALLDROP(制程) |
0.912 |
|
|
0.404 |
|
|
0.939 |
|
|
M1(制程) |
|
0.524 |
|
|
0.932 |
|
|
0.900 |
|
UBM(制程) |
|
|
0.838 |
|
|
0.888 |
|
|
0.880 |
MH(产品) |
0.921 |
0.490 |
0.820 |
|
|
|
|
|
|
RF(产品) |
|
|
|
0.468 |
0.965 |
0.916 |
|
|
|
TI(产品) |
|
|
|
|
|
|
0.949 |
0.818 |
0.826 |
All -
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
BALLDROP(制程) |
0.900 |
|
|
0.419 |
|
|
0.925 |
|
|
M1(制程) |
|
0.607 |
|
|
0.914 |
|
|
0.939 |
|
UBM(制程) |
|
|
0.769 |
|
|
0.864 |
|
|
0.547 |
MH(产品) |
0.926 |
0.565 |
0.844 |
|
|
|
|
|
|
RF(产品) |
|
|
|
0.487 |
0.936 |
0.910 |
|
|
|
TI(产品) |
|
|
|
|
|
|
0.934 |
0.742 |
0.290 |
All
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
All |
0.919 |
0.735 |
0.859 |
0.454 |
0.904 |
0.898 |
0.969 |
0.969 |
0.908 |
Single
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BallDrop |
0.918 |
0.485 |
0.829 |
0.458 |
0.765 |
0.876 |
0.936 |
0.824 |
0.938 |
MH-M1 |
0.880 |
0.656 |
0.661 |
0.357 |
0.889 |
0.850 |
0.903 |
0.717 |
0.412 |
MH-UBM |
0.630 |
0.204 |
0.874 |
0.123 |
0.477 |
0.838 |
0.746 |
0.687 |
0.713 |
RF-BALLDROP |
0.904 |
0.672 |
0.845 |
0.489 |
0.728 |
0.884 |
0.875 |
0.953 |
0.956 |
RF-M1 |
0.857 |
0.607 |
0.773 |
0.333 |
0.941 |
0.902 |
0.849 |
0.888 |
0.779 |
RF-UBM |
0.813 |
0.515 |
0.787 |
0.270 |
0.848 |
0.892 |
0.905 |
0.896 |
0.566 |
BALLDROP |
0.865 |
0.599 |
0.918 |
0.361 |
0.926 |
0.900 |
0.969 |
0.902 |
0.707 |
M1 |
0.853 |
0.595 |
0.720 |
0.338 |
0.913 |
0.891 |
0.817 |
0.973 |
0.877 |
UBM |
0.928 |
0.464 |
0.901 |
0.477 |
0.960 |
0.934 |
0.910 |
0.954 |
0.972 |
汇总结果
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
All |
0.919 |
0.735 |
0.859 |
0.454 |
0.904 |
0.898 |
0.969 |
0.969 |
0.908 |
MH-BallDrop |
0.918 |
0.485 |
0.829 |
0.458 |
0.765 |
0.876 |
0.936 |
0.824 |
0.938 |
MH-M1 |
0.880 |
0.656 |
0.661 |
0.357 |
0.889 |
0.850 |
0.903 |
0.717 |
0.412 |
MH-UBM |
0.630 |
0.204 |
0.874 |
0.123 |
0.477 |
0.838 |
0.746 |
0.687 |
0.713 |
RF-BALLDROP |
0.904 |
0.672 |
0.845 |
0.489 |
0.728 |
0.884 |
0.875 |
0.953 |
0.956 |
RF-M1 |
0.857 |
0.607 |
0.773 |
0.333 |
0.941 |
0.902 |
0.849 |
0.888 |
0.779 |
RF-UBM |
0.813 |
0.515 |
0.787 |
0.270 |
0.848 |
0.892 |
0.905 |
0.896 |
0.566 |
BALLDROP |
0.865 |
0.599 |
0.918 |
0.361 |
0.926 |
0.900 |
0.969 |
0.902 |
0.707 |
M1 |
0.853 |
0.595 |
0.720 |
0.338 |
0.913 |
0.891 |
0.817 |
0.973 |
0.877 |
UBM |
0.928 |
0.464 |
0.901 |
0.477 |
0.960 |
0.934 |
0.910 |
0.954 |
0.972 |
All-Product |
0.926 |
0.565 |
0.844 |
0.487 |
0.936 |
0.910 |
0.934 |
0.742 |
0.290 |
All-Process |
0.900 |
0.607 |
0.769 |
0.419 |
0.914 |
0.864 |
0.925 |
0.939 |
0.547 |
Uni-Product |
0.921 |
0.490 |
0.820 |
0.468 |
0.965 |
0.916 |
0.949 |
0.818 |
0.826 |
Uni-Process |
0.912 |
0.524 |
0.838 |
0.404 |
0.932 |
0.888 |
0.939 |
0.900 |
0.880 |
- 将各组模型在九种产品上的测试集表现进行归一化(与测试集相同源训练集模型的表现为1)结果:
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean |
MH-BallDrop |
1.000 |
0.739 |
0.949 |
0.937 |
0.813 |
0.982 |
0.966 |
0.847 |
0.965 |
0.911 |
MH-M1 |
0.959 |
1.000 |
0.756 |
0.730 |
0.945 |
0.953 |
0.932 |
0.737 |
0.424 |
0.826 |
MH-UBM |
0.686 |
0.311 |
1.000 |
0.252 |
0.507 |
0.939 |
0.770 |
0.706 |
0.734 |
0.656 |
RF-BALLDROP |
0.985 |
1.024 |
0.967 |
1.000 |
0.774 |
0.991 |
0.903 |
0.979 |
0.984 |
0.956 |
RF-M1 |
0.934 |
0.925 |
0.884 |
0.681 |
1.000 |
1.011 |
0.876 |
0.913 |
0.801 |
0.892 |
RF-UBM |
0.886 |
0.785 |
0.900 |
0.552 |
0.901 |
1.000 |
0.934 |
0.921 |
0.582 |
0.829 |
BALLDROP |
0.942 |
0.913 |
1.050 |
0.738 |
0.984 |
1.009 |
1.000 |
0.927 |
0.727 |
0.921 |
M1 |
0.929 |
0.907 |
0.824 |
0.691 |
0.970 |
0.999 |
0.843 |
1.000 |
0.902 |
0.896 |
UBM |
1.011 |
0.707 |
1.031 |
0.975 |
1.020 |
1.047 |
0.939 |
0.980 |
1.000 |
0.968 |
All |
1.001 |
1.120 |
0.982 |
0.929 |
0.960 |
1.007 |
1.000 |
0.996 |
0.934 |
0.992 |
All-Product |
1.008 |
0.861 |
0.966 |
0.996 |
0.995 |
1.021 |
0.964 |
0.763 |
0.298 |
0.875 |
All-Process |
0.981 |
0.925 |
0.880 |
0.857 |
0.971 |
0.968 |
0.955 |
0.965 |
0.562 |
0.896 |
Uni-Product |
1.003 |
0.746 |
0.938 |
0.957 |
1.025 |
1.027 |
0.979 |
0.840 |
0.849 |
0.929 |
Uni-Process |
0.993 |
0.799 |
0.958 |
0.825 |
0.991 |
0.995 |
0.969 |
0.925 |
0.906 |
0.929 |
- 将各组组合模型在九种产品上的测试集表现汇总成性能折线图:

相关结论
- 联合训练集有能力在测试集上展现出不差的表现,相较于使用与测试集匹配的训练集训练出的模型有比较稳定的差距;
- 多使用不同的数据训练 STPM 模型,确实可以一定程度上解决跨底板迁移困难的问题;
Hcsc 预训练网络 测试一
实验目的
- 测试 Hcsc 模型使用所有数据预训练得到的模型在 STPM 上的性能
实验设置
- Hcsc 网络预训练了 19 轮,没有达到论文相关要求,此处作为初步尝试的模型进行测试
- Hcsc 模型分为两种,由随机权重初始化的
hcsc_rand
和加载 ImageNet 预训练网络权重的 hcsc_pretrained
实验项目 |
实验设置 |
数据集 |
数据1 Total |
模型 |
Wide-Resnet50 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
Hcsc预训练数据 |
ImageNet,ADC dataset |
损失函数 |
L2 Loss |
网络预训练加载方式 |
教师加载 Base / Hcsc 学生加载 Base / Hcsc |
实验结果
- 采用在测试集上的二分类 AP 作为评判指标:
- T 表示教师网络类型,S 表示学生网络类型
NG AP |
weight_sum_1_3 |
G_mul_1_3 |
G_sum_1_3 |
G_weight_sum_1_3 |
L1_CC_L |
L1_CC_S |
L1_G_L |
L1_G_S |
L1_N |
L2_CC_L |
L2_CC_S |
L2_G_L |
L2_G_S |
L2_N |
L3_CC_L |
L3_CC_S |
L3_G_L |
L3_G_S |
L3_N |
L4_CC_L |
L4_CC_S |
L4_G_L |
L4_G_S |
L4_N |
lev_mul_1_3 |
lev_sum_1_3 |
Mul_CC_L |
Mul_CC_S |
Mul_G_L |
Mul_G_S |
Mul_N |
sigma_num_mul |
sigma_num_sum |
Sum_CC_L |
Sum_CC_S |
Sum_G_L |
Sum_G_S |
Sum_N |
weight_sum_1_3 |
T: hcsc_rand_19 S:imagenet |
0.897 |
0.893 |
0.915 |
0.916 |
0.873 |
0.895 |
0.889 |
0.902 |
0.847 |
0.904 |
0.916 |
0.913 |
0.916 |
0.873 |
0.868 |
0.872 |
0.870 |
0.873 |
0.857 |
0.887 |
0.883 |
0.873 |
0.885 |
0.853 |
0.866 |
0.871 |
0.909 |
0.915 |
0.901 |
0.913 |
0.882 |
0.782 |
0.772 |
0.899 |
0.899 |
0.901 |
0.902 |
0.875 |
0.872 |
T: imagenet S: hcsc_rand_19 |
0.860 |
0.856 |
0.867 |
0.868 |
0.829 |
0.839 |
0.843 |
0.849 |
0.795 |
0.860 |
0.868 |
0.866 |
0.867 |
0.834 |
0.859 |
0.855 |
0.857 |
0.853 |
0.840 |
0.857 |
0.854 |
0.849 |
0.854 |
0.840 |
0.832 |
0.832 |
0.868 |
0.869 |
0.861 |
0.868 |
0.843 |
0.701 |
0.744 |
0.867 |
0.862 |
0.863 |
0.863 |
0.849 |
0.835 |
T: hcsc_rand_19 S: hcsc_rand_19 |
0.895 |
0.891 |
0.914 |
0.914 |
0.868 |
0.893 |
0.886 |
0.900 |
0.840 |
0.903 |
0.916 |
0.913 |
0.919 |
0.870 |
0.899 |
0.907 |
0.899 |
0.910 |
0.861 |
0.888 |
0.890 |
0.820 |
0.893 |
0.738 |
0.865 |
0.867 |
0.915 |
0.923 |
0.887 |
0.914 |
0.864 |
0.782 |
0.771 |
0.911 |
0.919 |
0.890 |
0.923 |
0.780 |
0.868 |
T: imagenet S: hcsc_pretrained_19 |
0.887 |
0.890 |
0.912 |
0.913 |
0.869 |
0.879 |
0.881 |
0.890 |
0.824 |
0.897 |
0.913 |
0.910 |
0.917 |
0.857 |
0.894 |
0.901 |
0.895 |
0.903 |
0.854 |
0.859 |
0.871 |
0.842 |
0.869 |
0.782 |
0.854 |
0.857 |
0.906 |
0.913 |
0.890 |
0.907 |
0.866 |
0.808 |
0.799 |
0.902 |
0.906 |
0.901 |
0.907 |
0.864 |
0.859 |
T: hcsc_pretrained_19 S: hcsc_pretrained_19 |
0.877 |
0.883 |
0.906 |
0.906 |
0.876 |
0.898 |
0.896 |
0.904 |
0.838 |
0.887 |
0.907 |
0.906 |
0.911 |
0.853 |
0.891 |
0.904 |
0.893 |
0.906 |
0.849 |
0.884 |
0.881 |
0.889 |
0.886 |
0.740 |
0.856 |
0.853 |
0.906 |
0.918 |
0.888 |
0.909 |
0.868 |
0.784 |
0.770 |
0.903 |
0.915 |
0.906 |
0.918 |
0.801 |
0.854 |
T: hcsc_pretrained_19 S: imagenet |
0.878 |
0.885 |
0.907 |
0.908 |
0.876 |
0.898 |
0.897 |
0.904 |
0.837 |
0.889 |
0.908 |
0.907 |
0.911 |
0.853 |
0.887 |
0.893 |
0.888 |
0.896 |
0.830 |
0.848 |
0.861 |
0.846 |
0.864 |
0.813 |
0.855 |
0.854 |
0.903 |
0.911 |
0.884 |
0.907 |
0.858 |
0.796 |
0.775 |
0.878 |
0.895 |
0.878 |
0.900 |
0.824 |
0.856 |
T: imagenet S:imagenet |
0.885 |
0.889 |
0.914 |
0.916 |
0.866 |
0.879 |
0.880 |
0.893 |
0.825 |
0.897 |
0.916 |
0.912 |
0.922 |
0.855 |
0.899 |
0.909 |
0.899 |
0.911 |
0.844 |
0.890 |
0.899 |
0.867 |
0.902 |
0.790 |
0.851 |
0.855 |
0.906 |
0.920 |
0.885 |
0.909 |
0.859 |
0.808 |
0.809 |
0.902 |
0.915 |
0.903 |
0.917 |
0.837 |
0.857 |
T: hcsc_rand_200 S: hcsc_rand_200 |
0.900 |
0.905 |
0.922 |
0.923 |
0.876 |
0.904 |
0.901 |
0.911 |
0.840 |
0.912 |
0.920 |
0.922 |
0.926 |
0.874 |
0.906 |
0.913 |
0.907 |
0.917 |
0.877 |
0.835 |
0.836 |
0.770 |
0.845 |
0.723 |
0.875 |
0.872 |
0.906 |
0.915 |
0.890 |
0.910 |
0.868 |
0.811 |
0.795 |
0.897 |
0.908 |
0.893 |
0.914 |
0.828 |
0.874 |
- 采用高斯mask的第一~第四层特征作为获取结果的 map (L*_Gau_L)和
G_weight_sum_1_3
特征绘制折线图

H:Hcsc,R:rand,Pre:pretrained,Imgnet:Imagenet
相关结论
- 随机训练的 hcsc 作为教师时可以媲美 ImageNet 预训练模型的性能
- hcsc 以 ImageNet 权重为起点训练没有收益,甚至有性能下降
- 完全随机的模型训练后最深层特征性能较低
- 教师、学生网络预训练模型相同时整体性能较为稳定
当前结论以初步训练的 Hcsc 模型为基础,有可能完备训练的 Hcsc 模型性能更佳
Hcsc 预训练网络 测试二
实验目的
- 测试 Hcsc 模型使用
数据3
预训练得到的模型在 STPM 上的性能
实验设置
- Hcsc 网络预训练了 19 轮,没有达到论文相关要求,此处作为初步尝试的模型进行测试
- Hcsc 模型分为两种,由随机权重初始化的
hcsc_rand
和加载 ImageNet 预训练网络权重的 hcsc_pretrained
实验项目 |
实验设置 |
数据集 |
数据3 |
模型 |
Wide-Resnet50 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
Hcsc预训练数据 |
ImageNet,ADC dataset |
损失函数 |
L2 Loss |
网络预训练加载方式 |
教师、学生加载相同的网络权重 |
实验结果
hcsc 加载 ImageNet 网络
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean |
MH-BallDrop |
0.885 |
0.324 |
0.901 |
0.359 |
0.515 |
0.855 |
0.732 |
0.822 |
0.925 |
0.702 |
MH-M1 |
0.807 |
0.657 |
0.718 |
0.223 |
0.309 |
0.855 |
0.834 |
0.667 |
0.792 |
0.651 |
MH-UBM |
0.676 |
0.416 |
0.871 |
0.192 |
0.549 |
0.868 |
0.860 |
0.853 |
0.692 |
0.664 |
RF-BALLDROP |
0.889 |
0.363 |
0.914 |
0.424 |
0.469 |
0.826 |
0.716 |
0.798 |
0.934 |
0.704 |
RF-M1 |
0.773 |
0.466 |
0.921 |
0.209 |
0.842 |
0.861 |
0.776 |
0.964 |
0.885 |
0.744 |
RF-UBM |
0.594 |
0.482 |
0.761 |
0.151 |
0.510 |
0.897 |
0.863 |
0.796 |
0.542 |
0.622 |
BALLDROP |
0.794 |
0.594 |
0.961 |
0.272 |
0.475 |
0.844 |
0.953 |
0.927 |
0.962 |
0.754 |
M1 |
0.867 |
0.507 |
0.913 |
0.281 |
0.773 |
0.893 |
0.849 |
0.987 |
0.919 |
0.777 |
UBM |
0.890 |
0.461 |
0.927 |
0.395 |
0.771 |
0.898 |
0.883 |
0.915 |
0.976 |
0.791 |
All-Data |
0.900 |
0.687 |
0.941 |
0.357 |
0.855 |
0.867 |
0.880 |
0.982 |
0.956 |
0.825 |
- 采用在
数据3
9组测试集上的二分类 AP 的拐点分值作为阈值:
score |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
2.406 |
1.647 |
2.301 |
2.437 |
3.031 |
2.779 |
2.438 |
2.855 |
2.266 |
MH-M1 |
2.599 |
2.079 |
2.362 |
2.669 |
3.226 |
2.820 |
2.713 |
3.072 |
2.550 |
MH-UBM |
2.870 |
2.750 |
2.393 |
2.952 |
2.972 |
2.981 |
2.896 |
3.073 |
2.842 |
RF-BALLDROP |
2.289 |
1.778 |
2.206 |
2.203 |
2.957 |
2.807 |
2.450 |
2.860 |
2.322 |
RF-M1 |
2.577 |
2.711 |
2.305 |
2.814 |
2.529 |
2.892 |
2.618 |
2.608 |
2.529 |
RF-UBM |
2.854 |
2.586 |
2.118 |
2.911 |
2.962 |
2.466 |
2.663 |
2.742 |
2.754 |
BALLDROP |
2.707 |
2.687 |
2.164 |
2.773 |
3.078 |
2.968 |
2.635 |
2.879 |
2.326 |
M1 |
2.581 |
2.838 |
2.245 |
2.715 |
2.925 |
3.027 |
2.638 |
2.627 |
2.404 |
UBM |
2.682 |
2.995 |
2.039 |
2.770 |
2.940 |
2.987 |
3.008 |
2.930 |
1.821 |
All-Data |
2.261 |
2.032 |
1.882 |
2.309 |
2.338 |
2.516 |
2.294 |
2.272 |
1.969 |
hcsc 随机初始化网络 19 轮
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean |
MH-BallDrop |
0.905 |
0.258 |
0.911 |
0.348 |
0.739 |
0.870 |
0.888 |
0.715 |
0.904 |
0.727 |
MH-M1 |
0.812 |
0.704 |
0.706 |
0.352 |
0.630 |
0.859 |
0.919 |
0.729 |
0.708 |
0.713 |
MH-UBM |
0.641 |
0.350 |
0.612 |
0.191 |
0.726 |
0.779 |
0.860 |
0.716 |
0.370 |
0.583 |
RF-BALLDROP |
0.915 |
0.306 |
0.913 |
0.404 |
0.667 |
0.839 |
0.873 |
0.719 |
0.865 |
0.722 |
RF-M1 |
0.642 |
0.264 |
0.693 |
0.176 |
0.906 |
0.859 |
0.855 |
0.769 |
0.374 |
0.615 |
RF-UBM |
0.670 |
0.497 |
0.665 |
0.226 |
0.617 |
0.879 |
0.861 |
0.700 |
0.446 |
0.618 |
BALLDROP |
0.663 |
0.546 |
0.862 |
0.259 |
0.700 |
0.717 |
0.968 |
0.715 |
0.813 |
0.694 |
M1 |
0.785 |
0.413 |
0.795 |
0.282 |
0.902 |
0.883 |
0.882 |
0.851 |
0.537 |
0.703 |
UBM |
0.808 |
0.291 |
0.892 |
0.312 |
0.785 |
0.829 |
0.873 |
0.787 |
0.919 |
0.722 |
All-Data |
0.865 |
0.654 |
0.921 |
0.359 |
0.902 |
0.900 |
0.938 |
0.908 |
0.940 |
0.821 |
- 采用在
数据3
9组测试集上的二分类 AP 的拐点分值作为阈值:
score |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
2.320 |
2.004 |
1.801 |
2.613 |
2.911 |
2.696 |
2.814 |
2.968 |
2.060 |
MH-M1 |
2.438 |
2.037 |
1.825 |
2.565 |
2.935 |
2.507 |
2.576 |
2.836 |
2.125 |
MH-UBM |
2.761 |
2.340 |
1.527 |
2.698 |
2.891 |
2.742 |
2.734 |
2.947 |
1.837 |
RF-BALLDROP |
2.073 |
1.955 |
1.800 |
2.197 |
2.795 |
2.650 |
2.741 |
2.807 |
1.940 |
RF-M1 |
2.803 |
1.454 |
1.725 |
2.767 |
2.658 |
2.603 |
2.795 |
2.745 |
1.709 |
RF-UBM |
2.709 |
2.586 |
1.575 |
2.714 |
2.976 |
2.330 |
2.705 |
2.802 |
1.567 |
BALLDROP |
2.882 |
2.704 |
2.087 |
2.727 |
2.991 |
2.963 |
2.481 |
3.050 |
2.344 |
M1 |
2.697 |
2.987 |
2.067 |
2.734 |
2.859 |
2.611 |
2.932 |
2.785 |
2.695 |
UBM |
2.486 |
1.985 |
1.590 |
2.608 |
2.785 |
2.536 |
2.790 |
2.731 |
1.895 |
All-Data |
2.293 |
2.250 |
1.385 |
2.295 |
2.199 |
2.224 |
2.472 |
2.025 |
1.474 |
hcsc 随机初始化网络 200 轮
- 采用在
数据3
9组测试集上的二分类 AP 作为评判指标:
行为训练集,列为测试集
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean |
MH-BallDrop |
0.897 |
0.383 |
0.852 |
0.391 |
0.710 |
0.895 |
0.852 |
0.906 |
0.825 |
0.746 |
MH-M1 |
0.682 |
0.667 |
0.483 |
0.129 |
0.439 |
0.778 |
0.835 |
0.672 |
0.666 |
0.595 |
MH-UBM |
0.765 |
0.493 |
0.950 |
0.214 |
0.841 |
0.898 |
0.864 |
0.902 |
0.700 |
0.736 |
RF-BALLDROP |
0.876 |
0.323 |
0.831 |
0.454 |
0.606 |
0.857 |
0.783 |
0.737 |
0.856 |
0.702 |
RF-M1 |
0.726 |
0.467 |
0.869 |
0.149 |
0.980 |
0.850 |
0.761 |
0.967 |
0.632 |
0.711 |
RF-UBM |
0.655 |
0.474 |
0.787 |
0.137 |
0.635 |
0.922 |
0.786 |
0.857 |
0.310 |
0.618 |
BALLDROP |
0.801 |
0.407 |
0.933 |
0.315 |
0.686 |
0.861 |
0.957 |
0.931 |
0.923 |
0.757 |
M1 |
0.685 |
0.517 |
0.804 |
0.138 |
0.862 |
0.836 |
0.693 |
0.987 |
0.657 |
0.686 |
UBM |
0.859 |
0.135 |
0.741 |
0.358 |
0.765 |
0.887 |
0.869 |
0.778 |
0.976 |
0.708 |
All-Data |
0.884 |
0.662 |
0.931 |
0.412 |
0.936 |
0.898 |
0.943 |
0.969 |
0.942 |
0.842 |
ImageNet 预训练网络
- 采用在
数据3
9组测试集上的二分类 AP 作为评判指标:
行为训练集,列为测试集
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
Mean |
MH-BallDrop |
0.917 |
0.436 |
0.927 |
0.417 |
0.712 |
0.927 |
0.923 |
0.873 |
0.958 |
0.788 |
MH-M1 |
0.883 |
0.827 |
0.625 |
0.367 |
0.858 |
0.906 |
0.924 |
0.871 |
0.593 |
0.762 |
MH-UBM |
0.457 |
0.204 |
0.849 |
0.108 |
0.500 |
0.902 |
0.709 |
0.666 |
0.656 |
0.561 |
RF-BALLDROP |
0.902 |
0.453 |
0.868 |
0.430 |
0.774 |
0.916 |
0.879 |
0.943 |
0.963 |
0.792 |
RF-M1 |
0.659 |
0.591 |
0.885 |
0.222 |
0.979 |
0.916 |
0.874 |
0.883 |
0.796 |
0.756 |
RF-UBM |
0.710 |
0.445 |
0.810 |
0.193 |
0.686 |
0.932 |
0.905 |
0.853 |
0.608 |
0.683 |
BALLDROP |
0.864 |
0.647 |
0.950 |
0.345 |
0.887 |
0.943 |
0.966 |
0.956 |
0.847 |
0.823 |
M1 |
0.852 |
0.580 |
0.887 |
0.271 |
0.914 |
0.929 |
0.857 |
0.977 |
0.912 |
0.798 |
UBM |
0.916 |
0.441 |
0.919 |
0.429 |
0.924 |
0.946 |
0.935 |
0.961 |
0.971 |
0.827 |
All-Data |
0.911 |
0.758 |
0.907 |
0.398 |
0.969 |
0.937 |
0.952 |
0.975 |
0.950 |
0.862 |
- 采用在
数据3
9组测试集上的二分类 AP 的拐点分值作为阈值:
score |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
2.095 |
2.384 |
2.018 |
2.079 |
2.507 |
2.197 |
2.273 |
2.377 |
1.848 |
MH-M1 |
2.265 |
1.511 |
2.175 |
2.369 |
2.339 |
2.187 |
2.159 |
2.239 |
2.282 |
MH-UBM |
2.720 |
2.060 |
2.024 |
2.818 |
2.814 |
2.626 |
2.649 |
2.781 |
2.428 |
RF-BALLDROP |
1.948 |
2.306 |
1.969 |
1.652 |
2.307 |
2.067 |
2.192 |
2.221 |
1.836 |
RF-M1 |
2.365 |
2.209 |
2.046 |
2.292 |
1.960 |
2.243 |
2.334 |
2.308 |
2.193 |
RF-UBM |
2.370 |
2.211 |
1.884 |
2.370 |
2.284 |
1.979 |
2.089 |
2.221 |
2.240 |
BALLDROP |
2.423 |
2.301 |
2.164 |
2.324 |
2.384 |
2.297 |
1.977 |
2.446 |
2.149 |
M1 |
2.288 |
2.285 |
2.089 |
2.367 |
2.114 |
2.281 |
2.431 |
2.031 |
1.989 |
UBM |
2.416 |
2.607 |
1.982 |
2.381 |
2.510 |
2.646 |
2.433 |
2.454 |
1.600 |
All-Data |
1.972 |
1.646 |
1.742 |
1.874 |
1.874 |
1.864 |
2.011 |
2.056 |
1.548 |
汇总结果
- 使用不同预训练模型,在各个训练集训练出STPM 模型,将STPM模型在所有测试集上AP的均值绘制成表格

相关结论
- Hcsc 初步训练的模型用在 STPM 上,如果是随机初始化的模型开始训练,性能会比 ImageNet 预训练模型性能低百分之十以上
- 随机初始化训练 200 轮的 Hcsc 模型相比训练19轮的性能有了较大提升,和imagenet 预训练模型比低3-5个百分点,可以作为一种预训练模型的手段
- Hcsc 在 ImageNet 预训练权重基础上训练,总体来看可以提升模型性能
- 从阈值拐点结果中没有发现明显的具有结论性的规律
统计样本分值分布
实验目的
- 统计 STPM 模型在
数据3
上的二分类分值分布
- 寻找一种合适的阈值选取策略
实验设置
实验项目 |
实验设置 |
数据集 |
数据3 |
模型 |
Resnet50 |
Batch Size |
6 |
Epoch |
28 |
Init Learning Rate |
0.01 |
损失函数 |
L2 Loss |
实验结果
使用 9 组数据分别训练模型和综合所有训练数据训练模型,在 9 组测试集上统计分值分布情况
UBM |
 |
M1 |
 |
BallDrop |
 |
RF_UBM |
 |
RF_M1 |
 |
RF_BallDrop |
 |
MH_UBM |
 |
MH_M1 |
 |
MH_BallDrop |
 |
Total |
 |
相关结论
- 针对某一测试集性能较差的模型在 OK、NG 上的分值分布难以区分
- 选择阈值时可以统计所有测试数据,设置绝大部分OK 数据没有过杀的情况下的分值为阈值
- “大一统” 模型的阈值较为稳定 (实际项目中用的 1.7)
特征聚类分布
实验目的
- 探究特征维度上当前 BackBone 是否可以区分不同产品
- 跨底板迁移难度是否和特征分布相关
实验设置
实验项目 |
实验设置 |
数据集 |
数据3 |
模型 |
Resnet50 |
预训练模型 |
ImageNet |
特征层级 |
Layer1-4 |
实验结果
以 Resnet50 在 ImageNet 上的预训练模型为例。
Layer1
- 在 9类数据上的聚类:

- 在 9 类测试数据上的 NG AP 表现:
NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
0.673 |
0.302 |
0.860 |
0.173 |
0.222 |
0.782 |
0.747 |
0.793 |
0.838 |
MH-M1 |
0.702 |
0.851 |
0.484 |
0.150 |
0.138 |
0.831 |
0.573 |
0.729 |
0.764 |
MH-UBM |
0.220 |
0.114 |
0.784 |
0.029 |
0.111 |
0.579 |
0.332 |
0.643 |
0.728 |
RF-BALLDROP |
0.678 |
0.127 |
0.843 |
0.250 |
0.157 |
0.792 |
0.723 |
0.769 |
0.843 |
RF-M1 |
0.314 |
0.220 |
0.864 |
0.067 |
0.203 |
0.490 |
0.445 |
0.735 |
0.777 |
RF-UBM |
0.525 |
0.352 |
0.860 |
0.147 |
0.212 |
0.823 |
0.637 |
0.841 |
0.838 |
BALLDROP |
0.568 |
0.407 |
0.872 |
0.183 |
0.153 |
0.757 |
0.679 |
0.790 |
0.799 |
M1 |
0.506 |
0.290 |
0.874 |
0.116 |
0.220 |
0.645 |
0.491 |
0.818 |
0.811 |
UBM |
0.725 |
0.251 |
0.909 |
0.190 |
0.264 |
0.818 |
0.745 |
0.873 |
0.980 |
trace |
0.673 |
0.851 |
0.784 |
0.250 |
0.203 |
0.823 |
0.679 |
0.818 |
0.980 |
- 归一化(测试集、训练集匹配为1)后的 NG AP:
Norm_NG_AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
1.000 |
0.355 |
1.098 |
0.691 |
1.091 |
0.950 |
1.099 |
0.969 |
0.855 |
MH-M1 |
1.043 |
1.000 |
0.617 |
0.600 |
0.678 |
1.011 |
0.844 |
0.890 |
0.780 |
MH-UBM |
0.327 |
0.134 |
1.000 |
0.117 |
0.545 |
0.704 |
0.489 |
0.786 |
0.743 |
RF-BALLDROP |
1.007 |
0.149 |
1.075 |
1.000 |
0.773 |
0.962 |
1.064 |
0.940 |
0.860 |
RF-M1 |
0.466 |
0.258 |
1.102 |
0.268 |
1.000 |
0.596 |
0.655 |
0.898 |
0.793 |
RF-UBM |
0.780 |
0.413 |
1.097 |
0.588 |
1.044 |
1.000 |
0.937 |
1.028 |
0.856 |
BALLDROP |
0.844 |
0.478 |
1.113 |
0.732 |
0.752 |
0.920 |
1.000 |
0.966 |
0.816 |
M1 |
0.752 |
0.341 |
1.115 |
0.464 |
1.080 |
0.784 |
0.723 |
1.000 |
0.828 |
UBM |
1.078 |
0.295 |
1.159 |
0.762 |
1.297 |
0.994 |
1.096 |
1.067 |
1.000 |
distance |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
0.0 |
64.2 |
68.0 |
11.0 |
20.4 |
25.5 |
18.0 |
39.3 |
41.5 |
MH-M1 |
27.8 |
0.0 |
22.2 |
36.8 |
38.3 |
60.6 |
21.5 |
42.2 |
37.9 |
MH-UBM |
78.9 |
85.3 |
0.0 |
103.8 |
141.9 |
162.1 |
116.4 |
152.6 |
92.2 |
RF-BALLDROP |
12.8 |
57.5 |
77.3 |
0.0 |
19.7 |
26.8 |
14.3 |
38.5 |
43.3 |
RF-M1 |
45.3 |
82.9 |
125.7 |
30.8 |
0.0 |
18.3 |
25.7 |
18.9 |
55.9 |
RF-UBM |
43.4 |
92.7 |
125.8 |
27.5 |
10.8 |
0.0 |
31.5 |
20.0 |
57.7 |
BALLDROP |
52.8 |
42.5 |
107.5 |
23.7 |
24.8 |
38.4 |
0.0 |
37.1 |
64.0 |
M1 |
54.1 |
78.8 |
134.1 |
35.7 |
8.6 |
27.2 |
26.3 |
0.0 |
62.6 |
UBM |
27.3 |
44.1 |
22.0 |
36.7 |
26.5 |
29.7 |
26.3 |
19.2 |
0.0 |
- 各组训练集到每组训练集的距离与
1- Norm_NG_AP
之间的相关系数:
distance |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
corr |
0.79 |
0.77 |
-0.41 |
0.86 |
0.67 |
0.30 |
0.70 |
0.81 |
0.84 |
- 各组训练集到每组训练集的距离与
1- Norm_NG_AP
条形图:
 |
 |
MH-BALLDROP |
MH-M1 |
 |
 |
MH-UBM |
RF-BALLDROP |
 |
 |
RF-M1 |
RF-UBM |
 |
 |
BALLDROP |
M1 |
 |
|
UBM |
|
Layer2

NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
0.912 |
0.419 |
0.935 |
0.387 |
0.701 |
0.927 |
0.926 |
0.858 |
0.952 |
MH-M1 |
0.840 |
0.823 |
0.551 |
0.326 |
0.823 |
0.907 |
0.910 |
0.879 |
0.559 |
MH-UBM |
0.406 |
0.240 |
0.859 |
0.093 |
0.489 |
0.898 |
0.707 |
0.608 |
0.620 |
RF-BALLDROP |
0.898 |
0.425 |
0.878 |
0.397 |
0.768 |
0.923 |
0.911 |
0.929 |
0.956 |
RF-M1 |
0.623 |
0.592 |
0.899 |
0.211 |
0.966 |
0.913 |
0.869 |
0.882 |
0.772 |
RF-UBM |
0.670 |
0.425 |
0.833 |
0.175 |
0.604 |
0.934 |
0.903 |
0.852 |
0.559 |
BALLDROP |
0.833 |
0.648 |
0.943 |
0.294 |
0.826 |
0.941 |
0.957 |
0.934 |
0.833 |
M1 |
0.811 |
0.584 |
0.901 |
0.236 |
0.914 |
0.925 |
0.841 |
0.969 |
0.891 |
UBM |
0.899 |
0.426 |
0.926 |
0.414 |
0.877 |
0.947 |
0.929 |
0.944 |
0.970 |
- 归一化(测试集、训练集匹配为1)后的 NG AP:
Norm NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
1.000 |
0.510 |
1.089 |
0.975 |
0.726 |
0.992 |
0.967 |
0.885 |
0.981 |
MH-M1 |
0.921 |
1.000 |
0.641 |
0.820 |
0.852 |
0.971 |
0.951 |
0.907 |
0.576 |
MH-UBM |
0.445 |
0.292 |
1.000 |
0.233 |
0.506 |
0.962 |
0.738 |
0.628 |
0.639 |
RF-BALLDROP |
0.985 |
0.516 |
1.022 |
1.000 |
0.795 |
0.988 |
0.951 |
0.959 |
0.985 |
RF-M1 |
0.682 |
0.719 |
1.046 |
0.531 |
1.000 |
0.978 |
0.908 |
0.910 |
0.795 |
RF-UBM |
0.735 |
0.516 |
0.970 |
0.442 |
0.625 |
1.000 |
0.943 |
0.879 |
0.577 |
BALLDROP |
0.913 |
0.787 |
1.098 |
0.739 |
0.855 |
1.007 |
1.000 |
0.964 |
0.859 |
M1 |
0.889 |
0.709 |
1.049 |
0.595 |
0.946 |
0.991 |
0.878 |
1.000 |
0.918 |
UBM |
0.986 |
0.518 |
1.078 |
1.041 |
0.908 |
1.014 |
0.970 |
0.974 |
1.000 |
distance |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
0.0 |
49.5 |
115.2 |
6.8 |
21.2 |
20.7 |
19.9 |
27.3 |
75.0 |
MH-M1 |
28.9 |
0.0 |
27.1 |
28.7 |
52.2 |
59.8 |
18.1 |
42.6 |
20.9 |
MH-UBM |
95.6 |
64.6 |
0.0 |
101.5 |
145.7 |
140.9 |
77.6 |
134.7 |
63.5 |
RF-BALLDROP |
4.6 |
43.2 |
102.9 |
0.0 |
17.1 |
18.1 |
16.7 |
22.0 |
63.9 |
RF-M1 |
52.2 |
80.6 |
137.3 |
43.0 |
0.0 |
16.0 |
85.4 |
46.6 |
92.0 |
RF-UBM |
30.3 |
72.7 |
135.2 |
24.3 |
8.9 |
0.0 |
60.7 |
27.6 |
90.6 |
BALLDROP |
27.9 |
53.3 |
118.8 |
29.0 |
23.5 |
37.8 |
0.0 |
43.4 |
74.1 |
M1 |
36.8 |
74.7 |
145.4 |
37.2 |
12.4 |
25.6 |
41.4 |
0.0 |
95.4 |
UBM |
71.2 |
39.9 |
21.1 |
63.4 |
42.9 |
60.6 |
63.0 |
85.5 |
0.0 |
- 各组训练集到每组训练集的距离与
1- Norm_NG_AP
之间的相关系数:
distance |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
corr |
0.70 |
0.46 |
-0.41 |
0.58 |
0.64 |
0.50 |
0.60 |
0.76 |
0.06 |
Layer3

NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
0.804 |
0.347 |
0.764 |
0.334 |
0.706 |
0.823 |
0.813 |
0.561 |
0.792 |
MH-M1 |
0.655 |
0.540 |
0.721 |
0.281 |
0.730 |
0.692 |
0.842 |
0.442 |
0.472 |
MH-UBM |
0.576 |
0.158 |
0.690 |
0.164 |
0.578 |
0.761 |
0.773 |
0.406 |
0.354 |
RF-BALLDROP |
0.611 |
0.377 |
0.735 |
0.429 |
0.787 |
0.817 |
0.696 |
0.726 |
0.677 |
RF-M1 |
0.611 |
0.411 |
0.751 |
0.204 |
0.878 |
0.835 |
0.715 |
0.627 |
0.643 |
RF-UBM |
0.605 |
0.393 |
0.601 |
0.150 |
0.681 |
0.789 |
0.755 |
0.542 |
0.281 |
BALLDROP |
0.682 |
0.442 |
0.805 |
0.360 |
0.779 |
0.841 |
0.954 |
0.771 |
0.644 |
M1 |
0.620 |
0.408 |
0.691 |
0.269 |
0.800 |
0.861 |
0.763 |
0.873 |
0.655 |
UBM |
0.766 |
0.343 |
0.720 |
0.362 |
0.780 |
0.858 |
0.864 |
0.709 |
0.926 |
- 归一化(测试集、训练集匹配为1)后的 NG AP:
Norm NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
1.000 |
0.643 |
1.106 |
0.779 |
0.804 |
1.042 |
0.852 |
0.643 |
0.856 |
MH-M1 |
0.815 |
1.000 |
1.045 |
0.656 |
0.831 |
0.876 |
0.882 |
0.506 |
0.509 |
MH-UBM |
0.717 |
0.293 |
1.000 |
0.382 |
0.658 |
0.963 |
0.810 |
0.466 |
0.382 |
RF-BALLDROP |
0.760 |
0.699 |
1.065 |
1.000 |
0.896 |
1.034 |
0.729 |
0.831 |
0.730 |
RF-M1 |
0.760 |
0.762 |
1.089 |
0.477 |
1.000 |
1.057 |
0.749 |
0.718 |
0.694 |
RF-UBM |
0.753 |
0.729 |
0.871 |
0.349 |
0.776 |
1.000 |
0.791 |
0.620 |
0.303 |
BALLDROP |
0.848 |
0.820 |
1.166 |
0.839 |
0.888 |
1.066 |
1.000 |
0.883 |
0.695 |
M1 |
0.772 |
0.757 |
1.001 |
0.629 |
0.911 |
1.091 |
0.799 |
1.000 |
0.707 |
UBM |
0.953 |
0.635 |
1.043 |
0.845 |
0.888 |
1.087 |
0.906 |
0.812 |
1.000 |
distance |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
0.0 |
43.7 |
60.0 |
4.6 |
10.7 |
17.9 |
29.9 |
24.4 |
53.0 |
MH-M1 |
23.1 |
0.0 |
26.3 |
21.1 |
23.5 |
22.9 |
19.4 |
20.5 |
18.1 |
MH-UBM |
99.1 |
105.2 |
0.0 |
101.3 |
57.2 |
61.1 |
40.4 |
33.8 |
93.3 |
RF-BALLDROP |
2.1 |
39.0 |
59.3 |
0.0 |
10.0 |
18.0 |
25.1 |
23.3 |
48.4 |
RF-M1 |
27.5 |
49.7 |
51.2 |
29.0 |
0.0 |
6.7 |
24.8 |
19.9 |
53.2 |
RF-UBM |
25.5 |
45.9 |
51.3 |
27.2 |
3.7 |
0.0 |
23.7 |
19.1 |
46.7 |
BALLDROP |
66.6 |
73.4 |
35.2 |
66.1 |
31.3 |
44.7 |
0.0 |
12.5 |
76.7 |
M1 |
53.3 |
62.0 |
25.5 |
54.1 |
13.3 |
26.8 |
10.6 |
0.0 |
61.7 |
UBM |
50.4 |
31.2 |
28.7 |
45.0 |
37.4 |
30.0 |
16.2 |
30.9 |
0.0 |
- 各组训练集到每组训练集的距离与
1- Norm_NG_AP
之间的相关系数:
distance |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
corr |
0.31 |
0.74 |
-0.15 |
0.41 |
0.58 |
0.08 |
0.60 |
0.66 |
0.42 |
Layer4

NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
0.646 |
0.245 |
0.796 |
0.054 |
0.101 |
0.325 |
0.440 |
0.128 |
0.280 |
MH-M1 |
0.144 |
0.302 |
0.658 |
0.064 |
0.272 |
0.140 |
0.488 |
0.214 |
0.301 |
MH-UBM |
0.107 |
0.134 |
0.281 |
0.030 |
0.077 |
0.176 |
0.287 |
0.060 |
0.190 |
RF-BALLDROP |
0.118 |
0.222 |
0.738 |
0.147 |
0.236 |
0.434 |
0.522 |
0.370 |
0.136 |
RF-M1 |
0.105 |
0.269 |
0.700 |
0.057 |
0.645 |
0.278 |
0.568 |
0.256 |
0.619 |
RF-UBM |
0.251 |
0.252 |
0.367 |
0.060 |
0.102 |
0.367 |
0.647 |
0.125 |
0.078 |
BALLDROP |
0.116 |
0.330 |
0.746 |
0.085 |
0.222 |
0.269 |
0.849 |
0.190 |
0.321 |
M1 |
0.099 |
0.207 |
0.590 |
0.053 |
0.440 |
0.190 |
0.447 |
0.563 |
0.148 |
UBM |
0.148 |
0.176 |
0.620 |
0.070 |
0.130 |
0.226 |
0.533 |
0.136 |
0.127 |
- 归一化(测试集、训练集匹配为1)后的 NG AP:
Norm NG AP |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
1.000 |
0.812 |
2.832 |
0.369 |
0.156 |
0.887 |
0.518 |
0.227 |
2.199 |
MH-M1 |
0.223 |
1.000 |
2.341 |
0.438 |
0.422 |
0.381 |
0.574 |
0.381 |
2.365 |
MH-UBM |
0.166 |
0.443 |
1.000 |
0.201 |
0.120 |
0.481 |
0.338 |
0.106 |
1.496 |
RF-BALLDROP |
0.182 |
0.736 |
2.625 |
1.000 |
0.366 |
1.183 |
0.615 |
0.657 |
1.068 |
RF-M1 |
0.163 |
0.892 |
2.489 |
0.391 |
1.000 |
0.759 |
0.669 |
0.455 |
4.861 |
RF-UBM |
0.388 |
0.833 |
1.305 |
0.411 |
0.158 |
1.000 |
0.762 |
0.223 |
0.612 |
BALLDROP |
0.180 |
1.094 |
2.654 |
0.582 |
0.345 |
0.734 |
1.000 |
0.337 |
2.519 |
M1 |
0.154 |
0.686 |
2.101 |
0.359 |
0.682 |
0.517 |
0.526 |
1.000 |
1.166 |
UBM |
0.229 |
0.584 |
2.205 |
0.478 |
0.201 |
0.616 |
0.627 |
0.241 |
1.000 |
distance |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
MH-BALLDROP |
0.0 |
63.6 |
20.2 |
4.6 |
38.3 |
48.1 |
14.6 |
25.7 |
45.3 |
MH-M1 |
34.4 |
0.0 |
15.0 |
32.6 |
13.7 |
21.9 |
24.6 |
26.6 |
16.2 |
MH-UBM |
44.9 |
64.5 |
0.0 |
49.1 |
15.3 |
31.8 |
47.6 |
31.6 |
56.4 |
RF-BALLDROP |
2.4 |
56.2 |
12.8 |
0.0 |
27.8 |
39.2 |
14.5 |
23.5 |
41.2 |
RF-M1 |
40.5 |
40.2 |
10.8 |
44.7 |
0.0 |
15.6 |
41.9 |
20.7 |
38.9 |
RF-UBM |
38.0 |
50.1 |
10.7 |
42.5 |
9.3 |
0.0 |
42.0 |
20.6 |
34.7 |
BALLDROP |
40.3 |
93.5 |
42.6 |
41.5 |
52.6 |
45.9 |
0.0 |
14.6 |
83.5 |
M1 |
38.2 |
58.4 |
11.3 |
42.4 |
20.9 |
24.9 |
17.3 |
0.0 |
46.9 |
UBM |
35.2 |
36.1 |
31.2 |
40.1 |
39.7 |
28.1 |
70.4 |
56.6 |
0.0 |
- 各组训练集到每组训练集的距离与
1- Norm_NG_AP
之间的相关系数:
distance |
MH-BALLDROP |
MH-M1 |
MH-UBM |
RF-BALLDROP |
RF-M1 |
RF-UBM |
BALLDROP |
M1 |
UBM |
corr |
0.64 |
0.03 |
-0.57 |
0.64 |
0.45 |
-0.07 |
0.34 |
0.64 |
-0.18 |
相关结论
- 大部分数据集的距离和 NG AP 呈现出较明显的正相关,说明数据特征距离是影响跨底板迁移的一种反映手段
- 浅层特征的 AP 与 距离的相关性更高
- 但在实验中也出现了违法正相关结论的数据集 (M1),原因还不清楚,可能与降维方法与测试数据集选取有关
通道注意力
实验目的
实验设置
实验项目 |
实验设置 |
数据集 |
数据3 |
模型 |
Resnet50 |
预训练模型 |
ImageNet |
特征层级 |
Layer1-4 |

实验结果
以 Resnet50 在 ImageNet 上的预训练模型为例。
|
MH-BALLDROP |
MH-BALLDROP |
MH-M1 |
MH-M1 |
MH-UBM |
MH-UBM |
RF-BALLDROP |
RF-BALLDROP |
RF-M1 |
RF-M1 |
RF-UBM |
RF-UBM |
BALLDROP |
BALLDROP |
M1 |
M1 |
UBM |
UBM |
base(bad train date)1 |
0.918 |
0.951 |
0.785 |
0.852 |
0.931 |
0.958 |
0.398 |
0.775 |
0.945 |
0.953 |
0.948 |
0.978 |
0.951 |
0.959 |
0.979 |
0.997 |
0.960 |
0.977 |
base(bad train date)2 |
0.918 |
0.951 |
0.786 |
0.849 |
0.917 |
0.945 |
0.421 |
0.804 |
0.977 |
0.987 |
0.940 |
0.968 |
0.958 |
0.966 |
0.978 |
0.997 |
0.956 |
0.975 |
base (clean train data) |
0.915 |
0.949 |
0.779 |
0.843 |
0.921 |
0.947 |
0.415 |
0.801 |
0.971 |
0.980 |
0.939 |
0.968 |
0.958 |
0.966 |
0.979 |
0.996 |
0.958 |
0.978 |
cha-att |
0.839 |
0.866 |
0.740 |
0.797 |
0.924 |
0.951 |
0.360 |
0.692 |
0.937 |
0.942 |
0.951 |
0.973 |
0.951 |
0.960 |
0.974 |
0.987 |
0.929 |
0.946 |
cha-att-lr0.1 |
0.926 |
0.959 |
0.806 |
0.854 |
0.926 |
0.946 |
0.474 |
0.852 |
0.972 |
0.986 |
0.929 |
0.961 |
0.973 |
0.981 |
0.976 |
0.999 |
0.959 |
0.981 |
cha-att-lr0.1-remove top10% |
|
0.936 |
|
0.824 |
|
0.940 |
|
0.813 |
|
0.956 |
|
0.941 |
|
0.971 |
|
0.993 |
|
0.970 |
base(bad data lr0.1) |
0.919 |
0.955 |
0.748 |
0.784 |
0.905 |
0.921 |
0.459 |
0.836 |
0.969 |
0.979 |
0.923 |
0.953 |
0.974 |
0.982 |
0.959 |
0.978 |
0.960 |
0.977 |
base(clean data lr0.1) |
0.922 |
0.958 |
0.771 |
0.814 |
0.909 |
0.929 |
0.458 |
0.830 |
0.974 |
0.980 |
0.925 |
0.952 |
0.974 |
0.981 |
0.954 |
0.971 |
0.956 |
0.975 |
freeze cha att lr0.1 |
0.929 |
0.962 |
0.797 |
0.842 |
0.925 |
0.947 |
0.474 |
0.857 |
0.968 |
0.983 |
0.930 |
0.959 |
0.973 |
0.982 |
0.976 |
0.999 |
0.958 |
0.979 |
freeze cha att |
0.928 |
0.961 |
0.794 |
0.840 |
0.921 |
0.941 |
0.473 |
0.853 |
0.966 |
0.980 |
0.931 |
0.960 |
0.972 |
0.981 |
0.977 |
0.999 |
0.958 |
0.979 |
相关结论
- 在非跨底板迁移的情况下,大学习率的共享通道注意力具有更好的性能
- 注意力选择的前 10% 特征区分缺陷的能力较差
- 但剔除这部分特征后模型性能反倒下降
参考文献
[1] [Wang G , Han S , Ding E , et al. Student-Teacher Feature Pyramid Matching for Unsupervised Anomaly DetectionJ]. 2021.
文章链接:
https://www.zywvvd.com/notes/study/deep-learning/anomaly-detection/stpm-report/stpm-report/