Research On Industrial Product Surface Defect Detection Method Based On Deep Learnin

In the actual industrial production,defective products is inevitable limited by the existing technology and production environment,among which the most common is surface defects,such as scratches,wear,pores,fracture,etc.These defects may affect the performance,safety,life and aesthetics of the products,causing unpredictable impacts.Traditional manual inspection is difficult to meet the needs of high-speed production line operation,and has the disadvantages of low efficiency and high cost.With the rapid development of machine vision,surface defect detection technology is moving towards automation and intelligence,and product quality is further guaranteed.In the current defect detection research field,traditional machine vision-based methods are not widely available due to their low robustness,susceptibility to environmental interference and the presence of numerous manually adjusted parameters,making such methods only applicable to some specific scenarios.Although deep learning-based methods show their excellent detection performance,the fact that defective products are episodic in production makes fewer samples available for neural network training.In addition,the production line has certain requirements for real-time detection,and how to improve the speed as much as possible while ensuring the accuracy of model detection remains a major challenge in the field of surface defect detection.Current defect detection methods based on image semantic segmentation have become a popular research hotspot due to their good visibility,but most of these works only report the segmentation performance and can not provide a defect score for each image,ignoring that the decision of positive or negative is the core task of defect detection.To this end,this paper proposes two deep learning-based defect detection algorithms,which are implemented by designing segmentation network and decision network,it can be expected to solve the problem on small sample datasets with lightweight models.The specific work is as follows:(1)A four-stage surface defect detection method for industrial products is designed and implemented.In the first stage,the gray histogram stretching method is used to enhance the contrast between the foreground and the background,and the effect of illumination inconsistency is eliminated.In the second stage,a Mini U-Net is designed to segment the defect regions,which ensures the segmentation performance and simplifies the network.In the third stage,the segmentation results are corrected,some irrelevant interference is filtered through post-processing,and the defect candidate regions are output.In the fourth stage,a lightweight decision network is proposed,and the final defect score is obtained by combining the mask map output by the second stage segmentation network and the candidate defect area output by the third stage.The algorithm is performed on three public datasets: Magnetic Tile dataset,Kolektor SDD,DAGM.The results show that the model still has high detection accuracy and speed even when the defect samples are few.(2)An end-to-end surface defect detection network based on multi-scale adversarial mixed loss function is proposed.The network consists two parts: Generator and Discriminator.Generator(Segmentor)is improved lightweight Deep Lab V3+ network.In the adversarial training stage,a min-max game between the generator and the discriminator is played by introducing multi-scale perception loss function,which enables the generator to achieve good segmentation performance.In the classification training stage,defect detection is implemented by enabling classification head to transform discriminator into classifier.The algorithm is performed on public dataset Kolektor SDD2.The results show that the proposed model achieves excellent performance in segmentation performance and detection performance.The two algorithms in this thesis give consideration to both real-time requirement and accuracy.The lightweight network design makes them suitable for the scenarios where defect samples are scarce.Sufficient experiments show that they have the application value of deploying in industrial production lines.