Research On Metal Surface Defect Detection Technology Based On Weakly Supervised Learning

With the improvement of China’s manufacturing industry,the industrial products made in China are widely circulated all over the world because of their high quality and low price.Among them,the metal manufacturing industry is an important pillar to support China’s infrastructure and economic development,which is widely used in construction,transportation,electrical appliances and other fields.The rapid development of industry is inseparable from the output of high-quality metal industrial products.Due to some factors in processing and manufacturing,the products will inevitably have some surface defects.These surface defects not only affects the appearance of the products,but also has a certain impact on performance.In order to ensure the quality of metal products into the market and improve the competitiveness of enterprises,the surface defect detection of metal products has become a very important link in the process of industrial manufacturing.The initial surface defect detection mainly relies on manual identification,which is inefficient and affected by many subjective factors,and can not guarantee the high quality of products.The method based on machine vision needs to design features manually according to experience,which can not take into account a variety of defects detection.With the rapid development of deep learning,many visual detection methods based on deep learning have been successfully applied to the actual industrial production because of their far superior performance to traditional methods.However,the existing deep learning methods generally use the fully supervised training method,which requires a large number of data with instance level annotations.In some practical production areas,the acquisition of instance level annotations is expensive,while the acquisition of image level annotations is more convenient.In order to use the surface defect image without instance level annotations to realize the metal surface defect detection task,weakly supervised object detection technology is used to solve the problem of metal surface defect detection and realize the classification and location of surface defects(1)Object detection task is the combination of classification and localization task.In this paper,from the perspective of classification task,through the image classification experiment of the mainstream convolution neural network on the metal surface defect data set,the convolution network suitable for the metal surface defect detection task is selected as the feature extraction model in the current mainstream network.Aiming at the redundancy of the existing model learning features,the CRAM module is proposed to improve the feature representation ability of the existing network to the surface defect target.(2)By analyzing the existing weakly supervised object detection methods,convolution neural network is selected to generate feature activation maps to realize object localization.For the problem that CAM can only activate part of the area,this paper proposes the SGM method,which combines the deep and shallow features of the network to strengthen the network’s ability to locate the target,and carries out the relevant experimental verification.(3)According to the experimental results of defect location by localization network,data pseudo annotations are generated as supervision information,and then a fully supervised object detection model is trained to realize the metal surface defect detection task.At the same time,aiming at the lack of accuracy of the pseudo annotation,the positive and negative sample generation mechanism of the basic model is improved,and the appropriate loss function is selected to further improve the detection accuracy of the model...