A Real-time Metallic Surface Defect Detection Algorithm Based On YOLOX

During the production process,metal materials may develop defects on their surfaces due to manufacturing environment,equipment malfunctions,and operational errors.For example,aluminum profiles may develop orange peel,leakage,and pitting,while steel surfaces may develop cracks,punched holes,and indentations.These defects can significantly impact the reliability and safety of industrial products.To meet the demand for high-quality development of manufacturing industry in our country,it is crucial to detect defects on metal surfaces.This paper discusses a deep learning-based method for detecting metallic surface defects,aiming to design a high-precision,high-speed,and lightweight algorithm for real-time and accurate detection.This paper conducts in-depth research in two areas: object detection algorithms and data augmentation,with specific research findings as follows:To address the problems of poor generalization ability and low detection speed in existing metallic surface defect detection methods,a new detection algorithm,E-YOLOX,is proposed.This algorithm adopts a new feature extraction network ECMNet,which uses deep convolution to reduce network parameters;uses linear inverse bottleneck residual network to enhance feature extraction ability,and preserves more key features that are manifold distributed in low-dimensional subspaces within high-dimensional tensors during forward propagation;uses extended cross-stage partial network structure to diversify the gradient flow paths of neural networks,making deep neural networks learn and converge more efficiently;introduces implicit knowledge,effectively improving the detection effect;integrates spatial probability distribution information into the confidence branch,alleviating the problem of detection misalignment.Experimental results show that E-YOLOX-l achieves 48.3% m AP detection accuracy on steel surface defect dataset NEU-DET,37.1% m AP detection accuracy on steel surface defect dataset GC10-DET,and 78.4% m AP detection accuracy on aluminum profile surface defect dataset AL6-DET,respectively improving 1.6%,2% and 4.8%compared with baseline model YOLOX-l,while reducing parameter amount by 51%,computation amount by 47%,and achieving 54 FPS detection speed,increasing by 18 FPS.Compared with related algorithms,the new algorithm achieves higher detection accuracy and a better balance between accuracy and speed.Data augmentation can effectively improve the performance and robustness of object detection algorithms,and has important significance for improving the accuracy and stability of metallic surface defect detection.Aiming at the problems of difficulty in obtaining defect images and low quality of samples generated by existing data augmentation methods,new data augmentation methods Copy Paste and Edge Cutout are proposed.The Copy Paste method randomly adds defects on the training images,and restricts the categories of added defects to preserve the semantic information between defects and texture materials,and calculates the coverage ratio of added defects to ensure the integrity of defects.The Edge Cutout method generates masks and covers random regions of training images.It generates adaptive masks through pixels near image labels,making the network pay more attention to global features,enhancing edge features of defects,while avoiding generated masks being recognized as defects.Experimental results show that the two proposed data augmentation methods have improved the detection performance of E-YOLOX.The detection accuracy on the NEU-DET dataset reaches 49.6% m AP,on the GC10-DET dataset reaches 38.3% m AP,and on the AL6-DET dataset reaches 80.2% m AP,respectively improving by 1.3%,1.2% and 1.8%.