Research On Grain Boundary Defect Detection Algorithm Based On Deep Learning

Defect detection is an important part in materials research and application.Grain boundary defects are crystalline surface defects that cause the crystal structure to be susceptible to corrosion and significantly reduce the plasticity,hardness and tensile strength of the product.Currently,grain boundary defects are mainly detected and located by image processing techniques using images captured by High Resolution Transmission Electron Microscope(HRTEM).With the rapid development of deep learning technology,detection techniques based on deep neural network models have gradually become a hot research topic in the field of crystal materials.However,up to now there are relatively little research has been conducted on HRTEM images of grain boundary defects.In this paper,the characteristics of HRTEM images of grain boundary defects is investigated,and algorothms are proposed to improve the performance and speed of defect detection based on deep learning networks.The research content and innovation points are as follows.Firstly,a grain boundary defect detection algorithm based on improved Efficient Det is proposed to address the problems of high missed detection rate and low positioning accuracy in defect detection of grain boundary HRTEM images.Through a comparative analysis of the performance and speed of existing object detection networks,the Efficient Det network model is selected as the baseline for optimization.In order to reduce the missed detection rate,in the proposed algorithm the Efficient Det network is improved by introducing a feature fusion module to obtain weighted fusion of the feature layers extracted from the backbone network Efficient Net and Bi FPN(Bidirectional Feature Pyramid Network)network.By this means,the edge features with fine granularity of the grain boundaries is obtained.Furthermore,to improve the localization accuracy of defects,the localization loss function(L1 loss function)is replace with the CIo U(Complete Intersection over Union)loss function.The experimental results show that the improved algorithm has a better performance in detecting the positioning accuracy of grain boundaries than the original one,with an average improvement of 4.8% in AP(Average Precision)and5.4% in AR(Average Recall).Secondly,in order to speed up the algorithm and reduce the amount of parameters of the network,the pruning technique is investigated the model designed in previous section is compressed based on weight pruning and BN(Batch Normalization)layer channel pruning.By sorting the weithts of the learnable parameters in the convolutional layer and the fully connected layers,and zeroing the parameters with low weights,a sparse weight matrix is obtained.For the BN layer channel matrix,the L1 regularization is used to obtain the sparse solution.In order to maintain the performance of the compressed network,an iterative training process is performed to fine-tune the pruned parameters.Experimental results show that,for the compressed model,the number of parameters of the pruned model is reduced by 70%,the computation amount of FLOPs(Floating Point Operations)of the model is reduced by 60%,and the AP value is reduced by 2.4%.