Research On Surface Defect Detection Of Industrial Products Based On Multiple Exposure Enhancement And Retrospective Distillation Learning

With the progress of scientific research,our manufacturing and industry has been flourishing in the direction of high intelligence and automation since the founding of our country.Industrial surface defect detection as one of the important steps in the process of quality inspection of industrial products,and also as an important means of monitoring the production of industrial products.It is becoming a popular research problem in the field of computer vision.The difficulty of collecting industrial samples and the unknown nature of surface defects can lead to problems such as exposure anomalies and a wide variety of defects,which affect the performance of defect detection algorithms.To solve these problems,this thesis proposes a surface defect detection method for industrial products based on multi-exposure enhancement with retrospective distillation learning,and a prototype system for surface defect detection is designed and implemented.The main research work of this thesis includes the following:(1)This thesis proposes an industrial image enhancement method based on multi-exposure fusion.First,a linear interpolation method with histogram equalization is employed to acquire exposure-adjusted image sequences,and then the sequences are fed into a multi-scale embedding-based exposure fusion network to adaptively extract suitable image features from the image sequences to generate weight maps,while a feature fusion method based on spatial attention maps is introduced to be used to strengthen the feature extraction capability of the model and retain the original image details.MEF-SSIM is introduced for maintaining the consistency of the graph structure between the input and reconstructed images.Experimental results show that the method can significantly improve the quality of industrial images,and in turn improve the detection of defects indicated by industrial products.(2)A method based on retrospective distillation learning for industrial product surface defect detection is proposed.First,a knowledge distillation architecture trained in an unsupervised manner is used to solve the problem of unknown defect generation in real environments,while an intermediate layer feature extraction method is introduced to complement the teacher network feature information.Second,an iterative retrospective information fusion method is proposed to retrospectively deliver the feature outputs from the front layer of the teacher network for enhancing the discriminative ability of the student network for normal samples.In addition,a graph structure similarity metric is introduced to enhance the image space similarity of the distillation learning architecture.Finally,a gradient-based defect localization method is employed to obtain defect localization maps.The experimental results show that the method is better than the existing defect detection methods in most performance indicators.(3)A prototype system for intelligent industrial product surface defect detection is developed and implemented.The main system functional modules include industrial sample upload module,industrial image enhancement module,industrial surface defect detection module and result storage module.The experiments show that the system has good practicality and has high application value and prospect.