Research On Multi-task Driven Industrial Appearance Defect Recognition

As tasks combining computer vision and actual industrial production,defect detection and defect segmentation have broad research and application prospects in the fields of industrial consumer goods,highways and bridges,material parts and other fields.Defect detection presents industrial appearance defects in the form of rectangle boxes,defect segmentation presents pixel-by-pixel labels,and combining defect detection and defect segmentation,defect recognition can output both rectangular box labels and pixel-by-pixel labels of defects.The application of the general framework of joint detection and segmentation to defect recognition will cause a series of problems,insufficient information fusion of feature maps,too rough feature representation of defects,and competitive learning of defect detection and defect segmentation tasks.Therefore,this thesis conducts profoundly research on the limitations of the existing general framework,discusses the connection and differences between defect detection and defect segmentation,focuses on collaborative learning of defect detection and defect segmentation tasks,and constructs a high-performance multi-task industrial appearance defect recognition algorithm.The main work of this thesis is summarized as follows:(1)Aiming at the problem that the current general framework have low feature utilization efficiency and poor defect recognition performance,this thesis proposes a multi-task appearance defect recognition algorithm based on deep feature fusion and feature reconstruction.The feature fusion mode and the utilization mode of features in different tasks of Mask RCNN,a universal detection and segmentation joint framework,are analyzed in-depth,and two effective core modules are designed to improve the defect recognition performance.The first is the deep feature fusion module,which uses the bidirectional FPN structure and attention mechanism to fully fuse the information of different resolution feature maps.The second is the feature reconstruction module,which uses symmetric encoder-decoder structure and context aggregation to reconstruct the feature map of the segmentation branch.Experiments show that the proposed module effectively improves the defect recognition ability of the current framework,especially the defect segmentation performance is significantly improved.(2)In order to alleviate the competitive learning caused by defect detection and defect segmentation tasks sharing feature maps,this thesis proposes a feature interaction network to decouple the joint tasks of defect recognition into synergistic learning of defect detection and defect segmentation.The proposed feature interaction network allows sufficient channel mutual learning and position mutual learning between defect detection and defect segmentation feature maps through channel interaction module and position interaction module,which effectively improves the feature representation ability of feature maps and generates favorable and more targeted feature maps for subsequent tasks.(3)This thesis integrates the above two networks to propose a novel and high-performance multi-task appearance defect recognition algorithm for detecting and segmenting feature interaction.The proposed method obtains excellent defect recognition performance on both the appearance defect dataset of electronic devices and the Highway-crack dataset,and its performance not only far exceeds that of the benchmark network framework Mask RCNN,but also has stronger performance in defect segmentation than other leading segmentation methods commonly used in the field of defect segmentation.Specifically,the detection metrics such as AP and AR10 on the appearance defect dataset of electronic devices increased by 1.21 and 1.20 points compared with Mask RCNN,while the segmentation metrics such as m Io U,P,R and F1 increased by 5.24,7.45,2.25 and 5.0 points,respectively.