Recognition Of Metal Surface Defects Based On Deep Learning

The automatic detection of metal surface defects is a research hotspot in the field of industrial quality control.the high standard of metal surface quality imposed by industrial manufacturers put forward higher requirements for the performance of computer vision inspection system and its algorithm.Due to the complexity and diversity of metal surface defects,as well as the existence of interference such as texture areas,noise and dust,it is a challenge for traditional computer vision methods to detect scratches,cracks and dents.Traditional image processing needs to eliminate external interference for statistical analysis of defects and express specific defects in a qualitative or quantitative way,which has some limitations in application and performance.On the other hand,deep learning can independently learn more abstract highlevel representation in the data set,which can be robustly competent for the task of metal surface defect recognition.But often,the accuracy of deep learning model is accompanied by high computing costs and storage costs,which also limit the application of deep learning in resource-constrained embedded terminals.Therefore,this thesis studies a lightweight neural network for metal surface defect detection task,the main research work includes:1.In this thesis,a metal surface defect detection system based on traditional image processing method is proposed for four kinds defects including macula,scratch,crack and impurity in the research object.First of all,the electrical contact product in the collected image is divided into three parts: texture region,smooth region and outer circle region.Secondly,in order to reduce the influence of uneven illumination,the smooth region and the outer circle are divided into eight regions by eight fan-shaped masks.In different regions,different smoothness and gray mean are used as the discrimination threshold,in which the smoothness is mainly used to screen out scratch defects,and the gray average is mainly used to screen out macular defects.Finally,the morphological features are extracted by dynamic threshold segmentation to distinguish whether there are cracks and impurities.Through the layer-by-layer screening of the above three features,we can judge whether the product corresponding to the image to be tested is qualified or not,and the final accuracy is 91.1%.2.Compared with the traditional convolution neural network,this thesis uses transfer learning to construct a metal surface defect classification system based on MobileNetV2.The data set of the system is based on the pictures of metal products actually collected in the factory,and at the same time,in order to expand the amount of data,add the Northeastern University metal surface defect database,the final data set totals nine categories of defects and a normal product.By calculating the recognition rate of the network,the number of model parameters and the amount of model calculation,MobileNetV2 is compared with the other five network models,although the accuracy of MobileNetV2 is 0.10798% lower than that of VGG19,which has the highest accuracy.However,compared with VGG19,the number of parameters and the amount of computation are reduced by 98.4023% and98.4109%,respectively.By calculating the Accuracy,Precision Recall and F1-Score of the classification system based on pre-training MobileNetV2,it is proved that the proposed system has good and stable performance.Finally,the recognition rate of the metal surface defect classification system based on MobileNetV2 in the test set is 97.96%.3.On the basis of the above lightweight convolution neural network,the common methods of model compression are introduced,including pruning,quantization and low-rank decomposition and so on.In order to reduce the number of model parameters and the amount of model calculation,three different pruning methods L1-norm,Slimming and Auto Slimming are used to prune the metal surface defect classification system based on MobileNetV2 proposed in chapter 3.The experimental results show that the model has the best performance in the Slimming pruning method.The accuracy of the pruned system in metal surface defect recognition task is 95.89%,the size of the model is reduced from313.5M to 127.5M,and the number of model parameters is reduced from 2.24 M to 0.98 M.At the same time,the metal surface defect recognition systems of four different network models MobileNetV2,VGG,Res Net50 and Shuffle Net V2 under Slimming pruning are compared from the recognition rate before and after pruning,the change of the number of model parameters and the change of model calculation.The experimental results show that based on the pre-trained MobileNetV2 network model,under the action of Slimming pruning method,the accuracy of metal defect recognition is the highest,and the computational complexity and parameter complexity of the model are reduced to the lowest.It can realize the task of metal surface defect recognition on embedded and terminal devices,and provide some practical reference for the application of artificial intelligence in industrial production.