Research On Machine Vision Detection Method For Multi-type Surface Defects

At present,the development of deep learning provides many new technical solutions for industrial detection,which greatly improves detection accuracy and efficiency and is widely used in visual detection tasks in the fields of industrial product processing,medicine,and semiconductor.However,its application is very limited in the detection of various surface defects of industrial products due to three main reasons,that is,scarce defect samples and difficult sample collection,too many detection algorithm models and high detection costs,and many types of defects and changeable detection scenarios.Faced with these difficulties,the technologies of generative adversarial networks,target detection and model migration during deep learning were studied and improved in this dissertation,which solved the above three problems and experimentally verified by the detection of various surface defects of the motor commutator.The works done are mainly as follows:To solve the problem of scarce samples of defect images and difficult sample collection,a WGAN-based sample generation method,that is,Condition-based and Context-based Adaptive Wasserstein GAN（CCAW-GAN）,was proposed in this dissertation to generate defect samples and expand the defect data set.During the generation of the defect image,a Condition-based WGAN model was designed and an encoder-decoder generator with jump residual connection was proposed and designed,for to solve the problem that it is difficult to generate complex images in the WGAN model and improve the quality of defect generation.During the fusion of the defect and overall image,a Context-based WGAN model was designed and for the first time,a method of context generative adversarial training was proposed,aiming to make the fusion of the generated defect and the overall defect image more natural and to significantly simplify the generation process of defect image.In the meantime,the CCAW-GAN model was used to establish an extensive data set of commutator defect samples.To solve the problem of too many algorithm models and high detection costs in surface defect detection,in this dissertation,a defect detection method based on Fully Improved YOLOV4（FI-YOLOv4）was proposed,and an S-Res A unit was designed.Through the convolution and attentional mechanism of the multi-receptive field,the problems of the single scale and the indistinguishable importance of features in the YOLOv4 backbone network were solved.The B-SPP module was designed,and the maximum pooling and dilated convolution were mixed to solve the problem of the high similarity of SPP module features in YOLOv4.The cross-layer feature pyramid network was designed,intending to solve the problems of insufficient feature fusion and semantic information loss in the YOLOv4 feature pyramid module.The feature adaptive module was designed to solve the problem of single-scale feature prediction in YOLOv4,and the loss function Federal Focal Loss was designed to solve the problem of inconsistency between the screening index and the evaluation index of the detection framework in YOLOv4,comprehensively improving the performance of defect detection.By testing the defect detection data of the commutator,the FI-YOLOv4 model improves the target detection evaluation index AP by 8.6%compared with the original YOLOv4 model,effectively solving the problems of too many detection algorithm models and high detection costs in defect detection.To solve the problem of many types of defects and complex detection scenarios in the detection of surface defects,an improved Progressive Subnetwork Fusion Transfer Model（PSF-TM）algorithm was proposed for defect detection model migration,which solved the problem of low retention rate of knowledge and experience in the model migration.The methods of subnetwork sampling training and subnetwork integrated prediction were designed to solve the problem of local optimization in training during the process of model migration,and effectively prevent the phenomenon of model over-fitting.Compared with the conventional model parameter migration and frozen model parameter migration,PSF-TM improves the detection accuracy AP₅₀ by 27.6%and 21.9%,respectively,and the training time is only half of the original,effectively solving the problem of many types of commutator defects and complex detection scenarios.In addition,in this dissertation,the data set generation module,target detection module,and migration learning module were integrated.In view of the detection of surface defects of the motor commutator,an online web system was developed for intelligent visual detection of various surface defects,which includes the functions of model training,data visualization,data analysis,and real-time detection,and can realize the image generation of motor commutator surface defects,surface defect detection,and migration of various surface defect detection models.This verifies the effectiveness of the method proposed in this dissertation and provides a basis for the method studied in this dissertation to be better applied to the actual industrial scenarios.