Research On Fusion Model Of Optimized Convolutional Neural Network For Classification Of Hardware Image

Classification and retrieval of hardware is an important link in the production of many bag enterprises,therefore the research on a hardware image classification and retrieval system to meet the needs of actual production has important application value.This thesis is committed to solving the problems of poor accuracy of hardware image,unbalanced classification performance of various categories and the influence of image background.A hardware image classification algorithm based on fusion model of optimized convolutional neural network is proposed in this thesis.The main research content and work are as follows:(1)Aiming at the problems of poor accuracy,unbalanced classification performance of various categories and the influence of image background,a classification algorithm based on fusion model is proposed.Firstly,select the model(Efficient Net-B3)with the highest accuracy from the three types of networks,and then select the models(Res Ne Xt50 and Alex Net)with the higher accuracy from the remaining two types of networks,and the purpose is to select the models with the higher accuracy under various types of networks;Secondly,by comparing and analyzing the classification performance of the three models under various categories,the complementary model(Res Ne Xt50)to the classification performance of the Efficient Net-B3 model under various categories is selected,and the purpose is to balance the classification performance of various categories;Thirdly,the model output category probabilities of the Efficient Net-B3 model and the Res Ne Xt50 model are adopted the weighted average method in order to give different weights to obtain a fusion model with better classification performance;Finally,the hardware image is scaled and cropped,and the purpose is to highlight the classification target and reduce background interference,and the influence of image background is consequently alleviated.(2)The model optimization in fusion process.Firstly,the transfer learning is used to train the Efficient Net-B3 network and the Res Ne Xt50 network,and the parameters of the full connected layer are fine-tuned,and the purpose is to use the underlying features to improve the classification performance of the two models;Secondly,the image scale of Efficient Net-B3 network is altered,in order to compensate for the inherent suboptimal problem of the scaling network,thereby improving the classification performance of the model.(3)Verify the validity of the fusion model of optimized convolutional neural network.The proposed fusion model is tested and analyzed more fully on real hardware image data set,and the accuracy,precision,recall and F1-score of classification are 97.73%,96.41%,97.02% and 96.71.All these indicators are better than the algorithms of the comparison experiments,achieve the improvement in the classification performance of hardware image,which can better meet the actual application needs of bag enterprises.