Research On Steel Plate Surface Defect Processing Technology Based On Deep Learning

The detection of steel plate surface defects often completed by manual visual inspection and traditional machine vision in now.However,the detection accuracy of these methods will be changed due to the difficulty of extracting the essential features of defects and the interference of external environment.So,this paper will study a new detection method,which steel plate surface defect processing technology based on deep learning.In this paper,the VGG16 network model is introduced and improved.The improved VGG16 network model is used to detect the surface defects of steel plate,so as to solve the problems of traditional steel plate surface detection methods.The improvement method is as follows.Firstly,channel attention mechanism is added to the network model to improve the recognition accuracy.Secondly,the depthwise separable convolution is replaced by partial 3× 3 convolution in the network to reduce the amount of network parameters and calculation,and to increase the network nonlinearity.Finally,the Leaky Re Lu activation function is replaced by the Re Lu activation function in the network to solve the phenomenon of neuron death during the training process.The simulation results show that the improved method has a good detection effect on medium and large defects of steel plate in data set a,and the recognition accuracy can reach 98 percent,which is improved by about 14 percent.For small defects of steel plate surface in data set B,it needs to be improved,and the recognition accuracy is only 88 percent.Firstly,k-means method is used to cluster Anchor box to obtain more suitable size of steel plate defect Anchor frame.In this paper,we introduce and improve the YOLOv3 algorithm,and use it to detect small defects.The improvement method is as follows.Firstly,the Anchor box is clustered again by K-means method to obtain more suitable size of steel plate defect Anchor frame.Secondly,residual element is added to the shallow structure of YOLOv3 to obtain more information of shallow small targets.Thirdly,the prediction layer of 16 times down sampling and 32 times down sampling in the network model is discarded,and the features of 4 times down sampling and 8 times up sampling are fused to establish a new prediction layer of 4 times down sampling.Finally,the Inception-v3 is added to the Convolutional Set module to increase the depth and width of the network and improve the feature extraction ability of the deep network.The simulation results show that the average accuracy of the improved detection method is 94.1 percent,which is improved by about 19 percent.In this paper,the methods are proposed for the detection of large and small defects on the steel plate surface,which have the advantages of high accuracy,not easy to be disturbed by the outside world,and can effectively extract the essential characteristics of the steel plate surface defects,and it will be more widely used in industry.