因果推理引导的复杂花纹织物缺陷视觉检测深度学习方法  

Causal inference-guided deep learning method for vision-based defect detection of complex patterned fabrics

作　　者：梁天飚 刘天元 汪俊亮 张洁[1] LIANG TianBiao;LIU Tian Yuan;WANG JunLiang;ZHANG Jie(Institute of Artificial Intelligence,Donghua University,Shanghai 201620,China;College of Mechanical Engineering,Donghua University,Shanghai 201620,China;Department of Industrial and Systems Engineering,The Hong Kong Polytechnic University,Hong Kong 999077,China)

机构地区：[1]东华大学人工智能研究院,上海201620 [2]东华大学机械工程学院,上海201620 [3]香港理工大学工业与系统工程系,中国香港999077

出　　处：《中国科学：技术科学》2023年第7期1138-1149,共12页Scientia Sinica(Technologica)

基　　金：上海市教育发展基金会;上海市教育委员会“晨光计划”(编号:20CG41);国家工信部项目(编号:2021-0173-2-1);中国科协青年人才托举工程项目(编号:2021QNRC001)资助。

摘　　要：AI视觉质量检测是两化融合的先导场景,是复杂产品质量管控的重要手段.本文以复杂花纹织物为对象,提出因果推理引导的产品缺陷视觉检测深度学习方法,从而解决复杂背景干扰下的视觉检测难题.首先,构建复杂背景干扰下的缺陷检测结构因果模型,并提出阻断背景特征干扰的因果干预策略.其次,在因果干预策略的基础上建立缺陷特征敏感性神经网络(defect feature-sensitive neural network,DFSNN),包括两个特征提取模块(分别以同视角的无缺陷、有缺陷面料图像作为输入).然后,提出了因果关系敏感性学习模块,其差分两个特征提取模块的输出,并通过最大化输出差分来构建因果敏感损失函数,从而实现训练过程中对背景特征的阻断和对缺陷特征的敏感性学习.实验结果表明,DFSNN可有效减弱背景图案的混淆干扰,保持95%的缺陷识别准确率.Al-based visual quality inspection is a pioneering scenario for the integration of two technologies and an important tool for complex product quality control.The typical research product in this paper is complex patterned fabrics,and a causal inference-guided deep learning method for vision-based defect detection is proposed to solve the detection challenges under complex background interference.First,a structural causal model for defect detection under complex background interference is constructed,and a causal intervention strategy to block the confounding effects caused by the background feature is proposed.Second,a defect featuresensitive neural network(DFSNN)is established based on the causal intervention strategy,including two feature extraction modules that are fed with intact and defective fabric images from the same viewpoint,respectively.Then,during the training process,a causality-sensitive learning module is proposed,which differs the outputs of the two feature extraction modules and reconstructs the model loss function by maximizing the output difference so as to achieve the blocking of background features and the sensitivity learning of defective features.The experimental results show that DFSNN can effectively attenuate the confusion interference of background patterns and maintain 95%defect recognition accuracy.

关 键 词：织物缺陷检测 深度学习 因果推理 复杂背景 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TS107[自动化与计算机技术—计算机科学与技术]