Product Surface Defects Detection Method Based On Weak Supervised Segmentation

The surface defect detection of industrial products based on machine vision is an important part of industrial automation detection technology,which is very important to control product quality and enhance efficiency.With the successful application of AI technology represented by deep learning in recent years,defect detection algorithm based on deep learning has replaced the traditional algorithm based on artificial feature extraction step by step,due to its surpassing accuracy and good generalization performance.However,the widely used deep learning algorithm needs a lot of high-cost manual annotation work in the process of using,which seriously affects the efficiency of the algorithm.To reduce the workload of manual labeling,this paper studies the weakly-supervised surface defect detection algorithm of industrial products,focusing on solving the current situation that deep learning algorithm is highly dependent on manual labeling.This paper studies a weak supervised segmentation algorithm to achieve a stable and reliable defect detection effect relying on the image level label.A weak supervised segmentation algorithm based on CAM is proposed,which uses Class Activation Mapping(CAM)to roughly locate the defect area,and uses the siamese network as backbone to solve the affine transformation inconsistency.In view of characteristics of surface defects,a new pooling operation called global Log Sum Exp pooling is designed,and a parallel self-monitoring loss is proposed to locate the small defect area accurately.The downsampled-nonlocal module combined with shallow features is used to enhance the CAM.To verify the effectiveness of the algorithm proposed in this paper,we use the public product surface defect datasets for verification.The three improved modules proposed in this paper were verified through ablation experiments.In addition,we compare the algorithm proposed in this paper with the existing weakly-supervised segmentation algorithms,the experimental results show that our algorithm can be competent for the application scenarios of industrial surface defect detection,and get rid of the expensive manual labor of the existing algorithms based on deep learning.