Research On Surface Defect Detection Of Industrial Parts Based On Mask R-CNN

Surface defects on industrial parts,such as oil stains,spots,and scratches,can be detected through surface defect detection.This provides information on the defect category and location of the tested parts,which helps prevent unqualified products from entering the market.Deep-learning-based surface defect detection technology conforms to the modern trend of industrial development,and meets the practical requirements of industrial production.However,current feature extraction networks in existing models often overlook target defects that are small or located at the edge,making it difficult to identify parts with these defects and greatly reducing the accuracy of model detection.To address this issue,this article proposes an industrial part surface defect detection model based on the improved Mask R-CNN algorithm.The model adopts Res Net101 as its backbone network and improves FPN to effectively enhance the useful channel information of high-level feature maps.The structure of Mask R-CNN is further improved to achieve high-quality masks,resulting in an improvement in the accuracy of model detection.This study covers the following topics:(1)In this paper,we propose the S-Mask R-CNN model to enhance the feature extraction capability of the defect detection model.To address the challenge of overlooking image details during the feature extraction process,a bottom-up fusion pathway was added to the original FPN structure,enabling better utilization of shallow feature information.Additionally,we incorporated multi-region attention modules and differential region attention modules to capture defect location-sensitive feature maps and differences between adjacent feature maps.The fused differential features produce more informative detailed feature maps.Finally,we conducted ablation experiments and analyses on three different datasets of industrial parts defects,and the results demonstrate that the improved S-Mask R-CNN model achieves significantly better detection accuracy than other mainstream algorithms.(2)We propose the S+Mask R-CNN model to enhance the edge information extraction capability.To address the issue of overlooking image edge information during feature extraction,we introduce an encoder-decoder head and optimize the loss function based on the S-Mask R-CNN model.Firstly,we improve the mask branch by adding an encoder-decoder head to capture local-global information.Secondly,we refine the calculation of the Sobel operator for edge loss to obtain high-quality masks.Finally,we conduct experiments on three different defect datasets and compare them with other detection algorithms.The results show that the S+Mask R-CNN significantly improves the detection performance for different types of defects in various datasets.According to the experimental results,the proposed S+Mask R-CNN model exhibits high accuracy on three different defect datasets,achieving 96.8%,98.1%,and 94.7%,respectively.Compared with existing models,our model has significant advantages in effectively detecting surface defects of industrial parts and improving the pass rate of industrial products.Figure [39] Table [16] Reference [70]...