Research On Metal Plate Surface Defect Recognition Based On Convolution Neural Network

The metal industry is the foundation of China’s heavy industry and an important industry in the national economy.Surface defects of metal sheets not only affect the quality and aesthetics of downstream products,but also their defect detection technology represents the level of heavy industry in China.Due to the influence of the industrial production environment,the plate images collected by machine vision have problems such as high noise and uneven brightness.Therefore,tasks such as defect classification and segmentation of metal sheets are extremely challenging.At present,the surface defect recognition technology of metal sheet based on machine vision is one of the mainstream detection methods.Through machine vision,real-time,efficient and automated production tasks can be completed.On this basis,in view of the problem of insufficient classification accuracy of metal sheet defect samples,this paper uses traditional machine learning algorithms,combined with image processing technology of convolutional neural network(CNN),The identification of metal sheet surface defects is studied from the perspectives of image enhancement,image classification and image segmentation.The main contents are as follows:(1)For the imperfections of uneven illumination and gray level in the metal surface defect image,some specific algorithms were used to enhance the original image,such as HE,AHE,CLAHE,etc.In addition,the influence of illumination was removed by Retinex algorithm,which effectively reduced its influence for the image.Moreover,Gabor filtering was used in image feature extraction,which can remove some details that may affect the effect of defect detection,like stripes.In short,the improvement of the accuracy of image classification in the subsequent steps was realized by the fusion of a variety of image processing algorithms.(2)For traditional image classification problems,this paper uses the LBP feature encoding method,Convert the processed image to a one-dimensional feature vector,select 40-dimensional feature vectors and conducts classification research on traditional machine learning methods respectively.Then,using SVM as the main classification method,a comparative experiment of different image preprocessing algorithms was carried out.The results show that the recognition accuracy of the image after image preprocessing is improved by about 10% compared with the original image.Compared with the original image,the recognition accuracy of the filtered image is improved by about 4%.(3)For Convolutional Neural Network,Selected representative: Mobile Net,VGG16,Res Net50,etc,for the classification of defective images.Comparing image enhancement and SVM algorithm,the accuracy,precision and recall rate based on CNN defect classification are improved by about 10%.Then the backbone feature extraction networks such as mobilenet are selected to build the Mask R-CNN instance segmentation model,and the network model is optimized based on the attention mechanism of senet,CBAM and ECA.The results show that the optimized network is more stable in the segmentation effect of the defect area,and the m Io U reaches more than 84%.