Research On Defect Detection Algorithms Of Industrial Products Based On Weakly Supervised Learning

With the advent of the era of big data,deep learning is booming in the field of computer vision.However,the recognition effect of a strongly supervised learning target detection model often depends on a large amount of finely labeled data.In the industrial field,industrial products are iteratively updated quickly,and different products often have large differences in image features,and require high professional labeling.Therefore,compared with the natural field,the industrial field is often difficult to obtain a large amount of strong supervision and labeling data.In actual production,the industrial field still uses manpower recognition as the main method of image detection.This method is costly and also harms the health of workers.To solve this problem,this paper uses a target detection algorithm based on weakly supervised learning to detect the magnetic tile defect dataset.The defects in the industrial field are different from those in the natural field.The task of defect recognition has the characteristics of single background,unclear boundaries between defects and background,and many small defects.In response to these problems,this paper has made further improvements on the mainstream weakly supervised learning target detection model.The main research contents are as follows:(1)This paper proposes a weakly supervised target detection algorithm based on joint clustering.Since the prediction frame of the model in weakly supervised target detection is easy to locate the more obvious part of the target entity,in order to solve this problem,the candidate areas generated during the model training process can be clustered.For each candidate in the cluster This article introduces two labeling methods,one is to assign the same category label to the candidate areas in the cluster,the other is to treat all the candidate areas in the cluster as a package,and use the cluster label as the label of the package.In this paper,the two labeling methods are used at the same time to iteratively optimize the detector of the network model,which improves the positioning ability of the weakly supervised learning model.(2)Since the training data of strong supervised learning contains a positioning frame,the frame regression of the candidate area can be used to further improve the positioning effect.Based on the joint clustering improvement model,this paper uses the non-maximum value of the high score candidate area.Suppress to generate pseudo-labels as the regression target,so as to perform border regression on the candidate area.In this paper,the improved model and other mainstream weak supervision models are tested and compared on the magnetic tile defect data set.The results show that the improved model proposed in this paper has higher performance in small target detection,and the average index has increased by more than 16%.(3)This paper designs and implements an industrial defect detection software system,which integrates the detection algorithm into the visualization system.The staff can perform batch detection of industrial defects through the system,review the detection results,and output the detection report to improve The efficiency of the staff.