Research And Application On Surface Defects Recognition Of Sheet Mental Based On Deep Learning Research

Plates are one of the most important raw materials for industrial production,which is widely used in aerospace,machinery manufacturing and other fields.During the production and processing of plates,it is very easy to produce product quality problems,and the surface defects of plates are the most intuitive manifestation of product quality problems.At the same time,surface defects identification is an important part of industrial quality inspection,which is an important guarantee of product quality in the automated production process.Therefore,it is of great importance to study and apply the recognition of surface defects on plates.In this paper,based on the theoretical knowledge of deep learning and image recognition,we study and improve the recognition and classification algorithm for surface defects of plates:(1)To address the problem of varying size of defects on the surface of the sheet,a sheet surface defect recognition model based on multi-scale feature fusion is proposed.Firstly,drawing on the implementation of grouped convolution and deep separable convolution,four different sizes of convolutional kernels are used to construct a multi-scale feature fusion module to extract attention to different scale defect features;secondly,a hybrid attention mechanism is designed to focus on both channel and spatial features to improve the attention of the network to important features;then,the Selu activation function is used to replace the Relu activation function to solve the gradient disappearance and explosion problems.Finally,after experimental verification,the recognition accuracy of the proposed model in this paper can reach 99.6%,and the number of parameters is 28.39 M,which is a big improvement over the original ResNet50 and can pay attention to different scale defect features effectively.(2)To address the problems of noise interference in the plate surface defect images,the different shapes of different types of defects,and the difficulty of traditional convolutional neural networks to adapt to the geometric changes of defect targets,a plate surface defect recognition model based on residual shrinkage and deformable atrous convolution is proposed.Firstly,soft thresholding and attention mechanism are introduced into the ResNet50 network to establish a residual shrinkage ResNet50 network to enhance the recognition performance of the network under noisy conditions.Secondly,deformable convolution and atrous convolution are introduced into the residual shrinkage module to build the deformable residual shrinkage module and the atrous residual shrinkage module respectively to achieve effective attention to different defect shape regions and expand the feature extraction field;then,different residual shrinkage branches are built based on the above modules for feature fusion.Finally,experimental validation was conducted,and the model in this chapter improved 16.1%,15.6%,16.1%,and 16% in the four indexes of A,P,R,and F1,respectively,compared with the original ResNet50,which improved the accuracy of sheet surface defect recognition in a noisy environment and could better fit the irregularly shaped defects.(3)The system uses the Python language to write a sheet surface defect recognition management system,designs a GUI interface for human-computer interaction,and deploys the model training proposed in this paper into the system,realising the functions of user management,sheet management and defect recognition,reducing the difficulty of system operation and providing some specific application value.