Research On Deep Network Model For Product Surface Defect Detectio

In industrial production,surface defects in products are inevitable due to the manufacturing process and complex production environment.These defects not only affect the aesthetics of the product,but even affect the life and safety of the product.With the development of deep neural network technology,the automatic detection performance of defect areas has been effectively improved,and has become the mainstream product surface defect detection method at present.However,due to the rich variety of products in real production scenarios,complex backgrounds,noise interference and other degradation factors,current deep neural network-based product surface defect detection methods still suffer from the lack of defect boundary contour integrity,poor robustness and other problems.To address the above problems,this paper further delves into the pixel-level accurate segmentation network model for defect target detection,making full use of contextual feature information to improve boundary contour localization accuracy and designing a denoising module to enhance the robustness of the network to noise,with specific work including two aspects.(1)Defect detection network combining convolution and Swin-Transformer: In the actual product defect detection process,the defect areas are complex and diverse,while there are complex background interferences,making it difficult to detect the complete defect edges,and the accurate detection of defects needs to make full use of contextual information for discrimination.Therefore,this paper proposes a Swin-Transformer defect detection network that combines convolution.This paper first constructs the Swin-Transformer module incorporating convolution,and builds a multi-scale encoder-decoder network based on this module,designing an up-sampling layer in the encoder and a down-sampling layer in the decoder to give the Swin-Transformer encoder-decoder sequence features better detail information of local defect features.The multi-scale processing enables the network to learn explicit deep semantic information while fully extracting contextual information,enhancing the network’s ability to localize defect features and extract boundary contours.The network learning is further guided by a boundary structure loss function.Experimental results for multiple product types on two defect datasets show that the model designed in this paper can significantly improve the accuracy of defect detection and detect more complete defect edges.(2)Defect detection network based on multi-scale adaptive threshold shrinkage denoising:In actual production,in addition to the complex diversity of defect types and complex background interference,there is also the degrading effect of noise.Therefore,this paper proposes a multi-scale adaptive threshold shrinkage denoising defect detection network,combining soft threshold noise reduction with convolutional operations to design an adaptive threshold shrinkage denoising module that removes interference noise from features in each scale pathway and retains valid background information.In order to locate the defective object more accurately,a contextual 3D attention fusion module is designed to generate 3D attention maps by horizontal and vertical aggregation to enhance the anomalous region features.Finally,parallel multi-scale features are fused to achieve effective detection of different scales and different types of defects using depth supervision.Experimental results of the constructed model on two datasets and their noisy interference datasets are presented in this chapter to validate the superior performance of the method,and a prototype system for product surface defect detection is developed in conjunction with the algorithms in this chapter to present the defect detection results in a visualized form.