Research On Image Recognition Method Of Metal Fracture

Under the action of complex environment,the metallic materials in service will cause failure accidents such as fracture,corrosion and fatigue,which will cause significant economic losses and casualties.How to accurately and efficiently identify the metal fracture image and propose a new method to determine the mechanism of metal fracture is of great significance for the production and safe use of high-quality metal products.This paper combines artificial intelligence theory,introduces neural network method into image preprocessing,and introduces deep learning method into fracture type recognition,then a new image preprocessing and recognition method for metal fractures is given.The main research contents and results of the paper are as follows:1.An improved non-local means denoising algorithm based on neural network is designed.Combining the non-local means denoising algorithm with good denoising performance and retaining the complete characteristics of the boundary features,an improved non-local means denoising algorithm with four-layer neural network structure is designed.The improved algorithm jumps out of the idea of weighted average and exponential function,and avoids the effect of traditional non-local means denoising algorithm being influenced by artificial choice of the weight function and filter parameters,improve the single-layer fixed weight filtering structure in the original algorithm into a multi-layer weight updateable network structure.Experiments on visual quality,peak signal-to-noise ratio and structural similarity on the Barbara test chart verify the effectiveness of the proposed algorithm.Finally,an example of denoising the metal fracture image based on the algorithm is given.2.Design a metal fracture classify model based on convolutional neural network.The DMFC model construction process is as follows: production of metal fracture image datasets,determination of the number of model layers,determination of the size of the convolution kernel,the impact of the BN layer on the network performance,addition of the Dropout layer,and the overall structure of the DMFC model.The DMFC-A ～ DMFC-F models are designed respectively,and the training set,test set accuracy,error curve,and training time under various models are analyzed.Finally,the DMFC model is determined,and the accuracy comparison between the DMFC model and traditional machine learning methods KNN and SVM on the metal fracture image data set is given,which verifies the validity of the DMFC model.3.An image recognition method based on feature fusion of multi-perception regions of interest is designed.This part of the research is based on transfer learning,reinforcement learning and class activation map in deep learning.The VGG-16 network model,the VGG-19 network model and the residual neural network Res Net-50 model that have performed well in classification tasks in recent years are analyzed.According to the analysis conclusion: different models pay attention to different regions of interest when recognizing the same image.This paper presents three fusion schemes to fuse the models.Experiments show the accuracy comparison of three fusion models with the pre-fusion model and the DenseNet network model on the Kaggle dataset,verifying the effectiveness of the proposed fusion algorithm.Finally,an identification example of the method on the metal fracture dataset and a visual analysis of the class activation map are given.