Causality Guidance Learning Method For Generalizable Defect Detection Of Complex Patterned Fabrics

Complex patterned fabrics(CPF)are high-margin textile products,and their defect detection is of great significance for the textile enterprises to maintain the efficient quality control.Since the surface of CPF is covered with information-rich and structure-irregular background patterns,its defect detection needs to face engineering problems such as large information differences between background patterns and strong visual interference,which makes it a technical bottleneck for textile industry.In this paper,defect detection of CPF is studied.First,an unsupervised learning-based generalizable defect detection boundary analysis model for multi-pattern CPF is constructed.Second,a causality guidance learning method for defect detection of CPF and a defect-guided detection neural network are designed.The detailed research works are as follows:(1)Aiming at the collected CPF images with the characteristics that the patterns are with multitypes and the cost of image labeling is high,the analysis method of generalizable defect detection boundary of CPF is studied.According to the color and shape features of background patterns,a convolutional neural network-based fabric image feature encoder is constructed to realize the sampling and encoding of image features of multi-CPF;the feature measurement method is introduced to evaluate the similarity of feature embeddings to realize the division of detection boundaries of CPF with near distribution of image information.The experimental results show that the proposed method successfully realizes the acquisition of the feature information of the fabric background patterns without labels,and it is also adopted to divide nine kinds of CPF into two generalizable defect detection boundaries which can be seen as the data basis for the defect detection.(2)Aiming at the background patterns of CPF with characteristics of strong visual interference,the defect detection method of CPF under visual interference is studied.The structural causal model for defect detection of CPF is constructed and analyzed to obtain the causal intervention strategy for accurate defect detection;a defect-guided detection neural network including multi-scale feature difference mechanism is constructed to realize the feature-wise causal intervention of visual interference from background patterns;two-stage causality guidance learning method is proposed to realize the accurate defect recognition and detection.The experimental results show that the proposed method has achieved 94.04% and 93.95% detection accuracy on the data within two generalizable detection boundaries respectively,which is better than the current state-of-the-art fabric defect detection model.At the same time,the effectiveness of the proposed method for the learning of causal statistical correlation and the sensitivity extraction of defect features is proved through visualization experiments.(3)In order to meet the practical needs of the defect detection of CPF,a prototype system is built to provide an effective tool for engineering detection applications.The research results of this paper provide theoretical analysis and practical technology for the defect detection of CPF,and are of great value for textile enterprises to improve the efficiency of high-quality detection of CPF.