Research On Classification And Detection Of Surface Defects Of Industrial Parts Based On Deep Learning

In the field of industry,the surface quality of parts directly impacts equipment utilization and safety.Thus,manufacturing companies undertake rigorous quality inspections of their produced parts to ensure optimal appearance,corrosion resistance,wear resistance,and other critical properties.Failure to do so may lead to severe consequences.Unfortunately,manual and traditional visual inspection methods’ efficiency and accuracy are insufficient to meet enterprise quality control needs.In order to solve the above problems,this paper proposes the use of deep learning methods to construct algorithms for classifying and detecting surface defects in industrial parts.Furthermore,these algorithms are optimized by combining actual inspection objects.The main research focus of this paper is as follows:A surface defect classification method is proposed for strip steel surface classification,based on multi-scale feature fusion.First,the local features are aggregated using the VGG network as the Backbone.Then,convolutional kernels of different sizes are used to extract multi-scale features that are superimposed.An attention mechanism is then employed to capture more critical detail information.Finally,the features at different levels are fused to achieve strip steel surface defect classification.The proposed method achieves an accuracy of 99.86% on the NEU-CLS dataset with only 20% of defect samples as the training set,which addresses the challenge of collecting difficult sample data in industry.An improved YOLOv5 lightweight surface defect detection method is proposed and applied to surface detection of magnetic materials.Firstly,the SE attention mechanism,the C3 TR module embedded in Transformer Block and the SPPF module are introduced in Backbone to enhance the feature extraction performance of the Backbone network.Additionally,the number of parameters and model size are reduced by using C3 Ghost consisting of Ghost Bottleneck to replace part of the C3 module in the neck.Deep separable convolution is also applied and the original prediction branch in Head for detecting large targets is removed to speed up model training and detection.Experimental using the magnetic material dataset showed that improved YOLOv5 outperforms the original YOLOv5 model with an overall m AP improvement of 2.3% and a 7% improvement in small target detection accuracy.Moreover,the model reduces the number of parameters by 51.5%and the prediction weight model size by 50.1%,while increasing the FPS to 83.3,making the model easier to deploy and significantly more efficient to detect.A six-sided inspection prototype for detecting surface defects in magnetic materials is designed in response to the genuine requirements of a domestic manufacturer,based on actual scenarios.In light of the specific customer demands,optical experiments are carried out,camera lenses are carefully selected,mechanical structure schemes are designed,and an upper computer system is developed.After conducting testing,it is found that the prototype is able to effectively fulfill the customer’s inspection needs.