Research On Surface Defect Detection Algorithm Based On Small Samples

Deep learning methods are widely used in the detection of surface defects,especially in industrial products.However,collecting and labeling large amounts of defective samples is usually harsh and impracticable in some industrial products.This paper analyzes the problems encountered in the application of current defect detection algorithms in detail and proposes two defect detection algorithms,which are applied to small samples and non-defective samples.Defective sample acquisition is difficult,and the number of samples is difficult to meet the training of deep learning models.This paper proposes a small sample defect detection network algorithm(SSDN)based on the YOLO network,which optimizes the structure of the feature network and designs a lightweight feature extraction network.In the region proposal methods,the Anchor is improved to balance the accuracy and complexity of the algorithm detection.The SSDN algorithm is trained on the dataset DAGM2007 using only 10 images per sample.The Acc results are 100%,100%,99.20%,100%,99.29%,100%.The Ave-accuracy of the six class samples has reached 99.72%,and the accuracy is better than the 12-class CNN,SIFT,ANN,Weibull and 11-layer CNN algorithms.The types and characteristics of most surface defects are unpredictable,and the current unsupervised methods differ greatly in the detection of regular and irregular texture surface defects.This paper proposes a surface defect detection algorithm based on a fully convolutional autoencoder(Re Net-D).The Re Net-D includes two phases: reconstruction network training and surface defect detection.First,reconstruction network uses some non-defective samples to reconstruct original image,Second,image reconstruction residual maps are used as possible regions of defects,and the detection results are obtained through image processing algorithms.For irregularly textured surface defect samples,we propose SSIM-L1 as the loss function of the reconstructed network to improve the result of defect detection.The precision results of Re Net-D algorithm on DAGM2007,AITEX,Kylberg Sintorn Rotationdata dataset are 0.884,0.793,0.855,0.940,0.824,0.935.It is superior to the traditional unsupervised methods(LCA,PHOT)and deep learning-based unsupervised methods(MSCDAE).