Research On Surface Defect Detection Of Aluminum Profiles Based On Deep Learning

In the process of industrial production and transportation,defects on the surface of aluminum profiles will affect the appearance and quality of aluminum products.Therefore,it is necessary to detect defects on the surface of the aluminum profiles and sort out the defective products at the production line during the process.Traditional artificial recognition methods have low detection accuracy and efficiency.Detection methods based on traditional image processing are difficult to deal with complex and changeable aluminum profiles and complex defects.This thesis studies the classification and detection of surface defects of aluminum profiles based on deep learning methods.Then it improves the object detection methods in deep learning according to the characteristics of the aluminum profile surface data so that the defect detection algorithm can reach higher accuracy and robustness.This thesis introduces the research background of surface defect detection of metal materials such as aluminum profiles,and the current research status of related equipment and algorithms at home and abroad.The types and causes of aluminum profile surface defects is analysed.And the evaluation indicators of the aluminum profile surface defect detection algorithms used in this article is listed in detail.A method for surface defect detection of aluminum profiles based on classification and object detection networks is proposed.The classification methods based on ResNet and MobileNet and the detection methods based on Faster R-CNN,Cascade R-CNN and RetinaNet are introduced.Their advantages and disadvantages are compared and analyzed through experiments.An improvement based on feature extraction is proposed which aims at the problem that the detection network has difficulty in extracting subtle defects on the surface of aluminum profiles.It combines deformable convolution and feature extraction network to reduce the missed detection rate of subtle defects and improve network detection accuracy.An improvement based on the anchor boxes is proposed for the narrow and long defects,which caused the problem of missed detection.The K-means clustering algorithm and the Guided Anchoring anchor box automatic generation algorithm are applied to improve the anchor boxes' generation part of the network.It improves the detection rate of long and narrow defects,and reduces the overall missed detection rate of the model effectively.The detection network cannot train the non-defective images during training,which makes it easy to make false detections during detection.It is necessary to cascade the classification network and the detection network to perform effective defect detection.In response to this problem,improvements based on background augmentation are proposed.Data augmentation and Dual-Channel design are proposed respectively,so that non-defective images can also participate in the training process of the detection network.The false detection rate of the detection network is effectively reduced,and the classification accuracy and detection accuracy of the detection network are improved.Last,the work done in this thesis is summarized and the direction for further research is proposed.