Research On Classification And Detection Technology Of Aluminum Surface Defects Based On Deep Learning

During the production and transportation of aluminum profiles,a variety of surface defects will occur.These defects seriously affect the practicality and aesthetics of the product,and need to be classified and detected.In the field of machine vision,traditional image processing methods cannot cope with a wide variety of surface defects with complex backgrounds.In recent years,in the fields of computer vision and natural language processing,deep learning has shown far superior capabilities to traditional algorithms.Based on the method of deep learning,this article focuses on the classification and detection of aluminum profiles surface defects.The main contents include:In defect classification,this paper analyzed the types and causes of common aluminum profiles surface defects;the paper adopted transfer learning and data enhancement methods to solve the problem of small amount of surface aluminum profiles surface defect data;convolution kernel splitting and channel grouping were used to optimize the network structure.Finally,the paper completed the design of the aluminum profiles surface defect classification model based on the convolutional neural network.The classification model has achieved an average accuracy of 95%on the verification set.In defect detection,this paper improved the Faster RCNN model to deal with the difficulties encountered in aluminum profiles surface inspection,and designed an end-to-end aluminum profiles surface defect inspection model.Respectively,the data enhancement strategy was used to solve the problem of imbalanced defect samples;the improved backbone network enhanced the feature extraction capability of the model;the introduction of FPN solved the problem of large differences in sample size;the application of deformable convolution allowed irregular shape defect features to be better extracted;the introduction of soft-NMS methods reduced the occurrence of missed detections.Finally,the m AP of the detection model designed in this paper in the test set reached 82.85%,which is 20.55% higher than the original model,which proved the performance of the detection model and the effectiveness of the improvement strategy.