Research On Aluminum Surface Defect Detection Based On Target Detection Algorithm

Aluminum profile manufacturing is an important foundation for the development of China’s manufacturing industry.With the "One Belt And One Road" strategic deployment and promotion,the construction of transport infrastructure requires a large number of high-quality aluminum profiles,such as in power equipment,engineering machinery,railway equipment and other fields.With the development trend of global economic integration accelerating,high quality,short cycle and low cost have become three important factors for the survival and development of enterprises.The surface defect detection of aluminum profile is an important means of quality assurance.Nowadays,some enterprises in the detection link through the human eye to judge.However,the aluminum profile surface defects of many types,high similarity of defect types and small targets and other problems increase the difficulty of detection,which is difficult to meet the requirements of enterprise automation and intelligent management.With the rapid development of artificial intelligence,the detection system combining machine vision and neural network has become the mainstream program of enterprise quality inspection because of its stable and efficient performance.Therefore,this paper proposes an aluminum profile surface detection method based on deep learning to improve the detection accuracy of small target defects.The specific research contents are as follows:First of all,the current mainstream one-stage target detection and two-stage target detection are studied.According to the requirements of enterprises for high precision detection of aluminum profile surface defects,Faster RCNN based on two-stage target detection is determined as the basic network framework,and the network structure,candidate regional network,feature extraction network,multi-task loss function and other modules are analyzed.Soft NMS and ROI align are used to optimize the network,and replace the L1 loss function with CIOU to obtain more information of the detection box.Secondly,this paper uses clustering algorithm to improve the network accuracy.The size and number of anchor frames generated by the traditional RPN network are fixed,which fails to reflect the size distribution of defects in the actual data.K-means method is used to cluster aluminum surface defect data to redefine the anchor frame size.Through the experimental comparison,the GA-RPN network with clustering has higher detection accuracy.Thirdly,the adjustable mechanism is designed to solve the problem of excessive offset of anchor frame.In order to solve the problem of RPN network generating redundant anchor points,this paper uses GA-RPN network to generate candidate regions with high quality and low density.However,when the shape of the candidate frame is predicted by deformation convolution,the offset often exceeds the region of interest.Therefore,an adjustable mechanism is designed to correct the position information of the anchor frame.Finally,a multi-task FPN network structure is designed to improve the performance of small target detection.In view of the fact that most defect targets in aluminum are too small,which affects the detection accuracy and recall rate.On the basis of the traditional FPN network,a bottom-up feature fusion path is added to shorten the mapping distance between the high level features and the low level features.The weighted average sum of the last layer of the feature images of each stage in the feature extraction network is carried out to solve the inadequate fusion mode of sampling-feature fusion-resampling,and enhance the robustness of the network.