Research On Intelligent Recognition Model Of Copper Alloy Metallographic Spectrum Based On Deep Learning

With the development of modern science,high-performance copper alloys have been widely concerned and applied to the key industries.The properties and quality of copper alloys are affected by the composition of copper alloys.At present,the analysis of copper alloy composition is mainly completed by professionals through comparative analysis of microstructure images of copper alloy materials,so there are disadvantages such as high professional requirements,subjective factors of researchers,low efficiency and high cost.Therefore,it is essential to study the automatic identification technology of metallographic map of copper alloy materials.Aiming at the classification and recognition of metallographic maps of five types of copper alloys,this paper studies from the directions of data enhancement and convolutional neural network,and proposes three classification methods:(1)Three image enhancement methods are used to highlight the texture features of the microstructure,so as to improve the quality of the copper alloy metallographic samples.Afterwards,the size of the convolutional neural network feature extraction operator is adjusted and trained on the copper alloy metallographic dataset to obtain copper alloy metallographic recognition models and compare the recognition effects of each model under different processing methods.(2)Using the ECANet attention mechanism,fusion pooling and feature extraction operator,the magnification factor is explored and determined,the MobileNet V2 network is improved,and a lightweight convolutional neural network model suitable for microstructure identification of copper alloy is constructed.Finally,the effect of the improved lightweight network model on the recognition of metallographic images of copper alloys is analyzed,and the effect is compared with different convolutional neural network algorithms.(3)Combining maximum pooling and average pooling,using MBConv module and ECANet attention mechanism,deleting the intermediate redundant network layer,adjusting the width factor,and using MobileNet V2 as the main trunk network to improve,a lightweight network model with low energy consumption,high precision and low computational load is constructed.Finally,comparing and analyzing with various network models through different evaluation performance indicators.The experimental results show that the three metallographic identification methods proposed in this paper have high accuracy,low energy consumption and good classification effect.The main conclusions are as follows:(1)Through the training of the data augmentation samples,the fine-tuned four classical convolutional neural network models showed the best classification effect of ResNet34,reaching 96.51%,while the amount of fine-tuned ResNet34 network parameters was only1.69 M.The feasibility and applicability of the convolutional neural network model in the recognition of copper alloy metallographic diagram are verified.(2)The MobileNet V2 network model,which adds attention mechanism and integrates multi-scale feature information,has further improved and lightened the recognition accuracy of copper alloy metallography.The recognition accuracy of the model is 97.43%,the number of parameters is 1.39 M,and the calculation amount is 215.59 M.(3)The improved lightweight network model has been further improved in the recognition accuracy of copper alloy metallographic diagram,and a lightweight network model with low consumption,high efficiency and low complexity has been obtained.The identification accuracy is 97.51%,and the number of model parameters and calculation amount are 1.14 M and 76.81 M,respectively.