Lightweight Reconstruction Network for Surface Defect Detection Based on Texture Complexity Analysis

Deep learning networks have shown excellent performance in surface defect recognition and classification of certain industrial products. However, most industrial product defect samples are scarce and have a wide variety of defect types, making methods that require a large number of defect samples for training unsuitable. In this paper, a lightweight surface defect detection network (LRN-L) based on texture complexity analysis is proposed. Only a large number of defect-free samples, which can be easily obtained, are needed to detect defects. LRN-L includes two stages: texture reconstruction stage and defect localization stage. In the texture reconstruction phase, a lightweight reconstruction network (LRN) based on convolutional autoencoder is designed, which can reconstruct defect-free texture images; a loss function combining structural loss and L1 loss is proposed to improve the detection effect; we built a calculation model for image complexity, calculated the texture complexity for texture samples, and divided textures into three levels based on complexity. In the defect localization stage, the residual between the reconstructed image and the original image is taken as the possible region of the defect, and the defect localization is realized via a segmentation algorithm. In this paper, the network structure, loss function, texture complexity and other factors of LRN-L are analyzed in detail and compared with other similar algorithms on multiple texture datasets. The results show that LRN-L has strong robustness, accuracy and generalization ability, and is more suitable for industrial online detection.