Self-supervised Industrial Unknown Defect Detection Based On Template Enhancement

The detection of surface defects of industrial products is a crucial link in the industrial production process.The quality monitoring of industrial products can be achieved through efficient detection methods of industrial products defects,which can provide a strong guarantee for the beauty,comfort and performance of products.Traditional defect detection methods based on machine vision have high requirements for data sets and great limitations,and poor adaptability.In recent years,with the development of convolution neural network theory and technology,the deep learning method has gradually become the mainstream to realize the defect detection of industrial products.In this thesis,aiming at the detection to the unknown defects of the industrial products,the deep learning based unknown defect detection algorithms are researched.The main work is as follows:In order to solve the problems such as the small number of defect data sets,the interference of background environment,and the poor reconstruction effect,a defect detection method based on self-supervised mode is proposed.Industrial product defect detection also faces a large number of unlabeled data set samples.The number of defect samples is far less than the number of defect-free samples.The self-supervised method can only apply defectfree samples for training,reducing the requirements for data sets.The background environment interference is another problem that needs special attention in industrial defect detection.Due to the complexity of background factors and the characteristics that are difficult to eliminate,the detection accuracy of defects will be significantly affected.To solve this problem,the defect-free image is reconstructed by using the generative method,and the template information is integrated into the information distribution of the image to be detected by introducing the attention module in vision transformer(Vi T).These measures improve the quality of the reconstructed image and are conducive to the detection of bottle cap defects.The proposed detection method is verified by experiments.The experimental results show that the new method improves the quality of the reconstructed defect-free image,and the recall rate increases by about 19% compared with the original UAE method.Research on the method of introduction template fusion is conducted.During training and testing,if the model can focus on the defect location,the defect detection rate can be improved.The defect location can be provided by the corresponding template information,and the rest location information is determined by the distribution information of the input image.The input defect image is taken as the main line of the network model,and part of the effective distribution information of the template is fused into the information distribution of the input image by an auxiliary method.The quality of the reconstructed defect-free image is improved by the combination of primary and secondary network structures,soas to improve the defect detection rate.The test under the original UAE framework shows that the new fusion method has significantly improved in three indexes,indicating the effectiveness of the new fusion method.