基于梯度引导加权‒延迟负梯度衰减损失的长尾图像缺陷检测  

Long-tailed image defect detection based on gradient-guide weighted-deferred negative gradient decay loss

作　　者：李巍 梁斯昕 张建州[1] LI Wei;LIANG Sixin;ZHANG Jianzhou(College of Computer Science,Sichuan University,Chengdu Sichuan 610065,China)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《计算机应用》2023年第10期3267-3274,共8页journal of Computer Applications

基　　金：中国博士后科学基金资助项目(2022M712235);四川大学专职博士后研发基金资助项目(2022SCU12074)。

摘　　要：针对目前图像缺陷检测模型对长尾缺陷数据集中尾部类检测效果较差的问题,提出一个基于梯度引导加权‒延迟负梯度衰减损失(GGW-DND Loss)。首先,根据检测器分类节点的累积梯度比值分别对正负梯度重新加权,减轻尾部类分类器的受抑制状态;其次,当模型优化到一定阶段时,直接降低每个节点产生的负梯度,以增强尾部类分类器的泛化能力。实验结果表明,在自制图像缺陷数据集和NEU-DET(NEU surface defect database for Defect Detection Task)上,所提损失的尾部类平均精度均值(mAP)优于二分类交叉熵损失(BCE Loss),分别提高了32.02和7.40个百分点;与EQL v2(EQualization Loss v2)相比,分别提高了2.20和0.82个百分点,验证了所提损失能有效提升网络对尾部类的检测性能。Aiming at the problem that the current image defect detection models have poor detection effect on tail categories in long-tail defect datasets,a GGW-DND Loss(Gradient-Guide Weighted-Deferred Negative Gradient decay Loss)was proposed.First,the positive and negative gradients were re-weighted according to the cumulative gradient ratio of the classification nodes in the detector in order to reduce the suppressed state of tail classifier.Then,once the model was optimized to a certain stage,the negative gradient generated by each node was sharply reduced to enhance the generalization ability of the tail classifier.Experimental results on the self-made image defect dataset and NEU-DET(NEU surface defect database for Defect dEtection Task)show that the mean Average Precision(mAP)for tail categories of the proposed loss is better than that of Binary Cross Entropy Loss(BCE Loss),the former is increased by 32.02 and 7.40 percentage points respectively,and compared with EQL v2(EQualization Loss v2),the proposed loss has the mAP increased by 2.20 and 0.82 percentage points respectively,verifying that the proposed loss can effectively improve the detection performance of the network for tail categories.

关 键 词：长尾数据集 累计梯度比值 加权损失 图像缺陷检测 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]