Surface Defect Detection Of Automobile Wheel Hubs Based On Semantic Segmentation

Surface defect detection of automobile wheel hubs is important to automobile quality control.As an accurate solution,the surface defect detection technology based on computer vision has been extensively studied and gradually put into actual production.Recently,deep learning has been widely used in computer vision,such as classification,object detection,and semantic segmentation.In order to further improve the accuracy of defect detection,how to use deep learning to detect surface defects is one of the current research hotspots.However,there are still few researches on the surface defect detection of automobile wheel hubs.The main reasons are the difficulty of producing automobile wheel hub defect datasets,the large variation of defect scales,the various defect types,and the difficulty in improving the effect of defect detection.In this paper,we study the surface defect detection network of automobile wheel hubs based on semantic segmentation,which can accurately obtain the position of the defect area at the pixel level from the automobile wheel hub image.we propose some specific improvement methods for the characteristics and difficulties of defect detection:Firstly,a deeper convolutional neural network can improve the accuracy of classification,but it will perform multiple downsampling,which greatly reduces the spatial resolution of the feature and cannot accurately predict the boundary and contour of the defect area.In this paper,High-Resoultion Net(HRNet)is used as the backbone network to perform the fusion of different resolution features when extracting features,which can extract effective high-resolution features.Secondly,due to the different sizes of defects,if a single-scale convolution kernel is used to extract features,defects with too large or too small size cannot be well detected.In this paper,a feature pyramid structure based on atrous convolution is used to extract features,which makes it possible to extract the multi-scale features to improve detection accuracy.At the same time,in order to solve the local information loss caused by the grid effect of atrous convolution,the layered superposition processing is carried out for different scale features in this paper.Then,this paper designs a decoder which can fuse the multi-scale feature with high-resolution feature to obtain fine segmentation results.Thirdly,since the appearance of the defect edge area is very similar to the surrounding normal area,the detection accuracy in the defect edge area will be very low.In this paper,based on the optical flow method,the body features of the image are obtained,and then the edge features are obtained by subtracting multi-scale feature from the body feature.At last,the supervised training on body feature and edge feature are carried out simultaneously to improve the detection accuracy of the edge area.In this paper,an experimental comparison on the automobile wheel hub defect dataset and the other three industrial product defect detection datasets is conducted.The experimental results prove that the proposed method in this paper can obviously improve the accuracy of automobile wheel hub defect detection and is suitable for a variety of industrial product defect detection scenarios.