Multi-feature Fusion Target Retrieval Algorithm Based On Deep Learning

The target retrieval technology has important practical research value,and the most important part is the construction and selection of the feature.The traditional methods use manual feature extraction,more people to participate in,the feature expression ability is poor,retrieval effect is poor.The abstract features of images can be automatically extracted by using deep learning technology,and the ability to express the features is greatly enhanced.A target retrieval algorithm based on deep learning and multi-feature fusion is proposed in this paper.Face image is taken as the object of retrieval to study the retrieval performance of algorithm.Firstly,the influence of parameters on retrieval results is studied,and the experimental parameters are determined by analyzing the results.Secondly,wavelet feature,tensor PCA feature(TPCA)and convolution neural network(CNN)are extracted respectively.Finally,the weighted fusion algorithm of traditional features and CNN features is studied.The results show that the accuracy of the proposed algorithm is higher than that of the conventional algorithm,and the specific work is as follows:(1)Research the influence of the parameters on the retrieval results.Using Db2,Db4,Bior2.4,Coif2 and Sym2 wavelets,the number of decomposition layers from 1 layers to 3layers is studied in order to study the effect of wavelet basis function and decomposition level on retrieval results.The results of the experiment are analyzed,The wavelet function of ORL library is selected as Db4 and the number of decomposition layers is 3,the best retrieval effect is 96.25%.The effects of CNN model parameters on the number of full connected layer neurons,the number of convolution layer features,the size of the convolution layer,the learning rate,the batch block,the convolution kernel and the sampling size have been studied.The experimental results are analyzed,When the number of full connected layer neurons in the ORL library is100,the number of convolution layer features is respectively 10,15,the image size is56?56selected,the learning rate is 0.35,and the batch block size is 40,When the convolution kernel and sampling size are 5?5 and 2?2,the best retrieval result is 97.5%.When the number of full connected layer neurons in the MIT face library is 200,the number of convolution layer features is 20,30 in turn,the image size is 64?64,the learning rate is 0.35,the batch size is 50,the convolution kernel and the sampling size are 5?5 and 2?2,the best effect of the image retrieval is 95%.(2)Research the strategy of feature fusion.Firstly,we study the contribution of wavelet components to retrieval results,and achieve the fusion of wavelet features.Secondly,we studythe retrieval accuracy of wavelet fusion features,tensor PCA features and deep CNN features respectively.Then,the fusion methods of the wavelet feature,the TPCA feature and the CNN feature are determined to determine the fusion coefficient.and the comparison of Squared Chi-Squared distance,Cosine distance,Euclidean distance and Manhattan distance is studied to determine the best similarity measure function of the classification.The experimental analysis shows that when the weighted fusion coefficient of the feature is 0.1,0.2,and 0.7 on the ORL library,the retrieval effect is best when the Manhattan distance is used,and the average accuracy is up to 98.75%.(3)The performance verification and analysis of the paper algorithm.On the ORL library,the CNN deep feature retrieval algorithm,the wavelet feature and the TPCA feature fusion algorithm,the wavelet feature and the PCA feature fusion algorithm and the proposed algorithm are compared.The experimental results show that the retrieval rate of this algorithm is 1.75%higher.The retrieval algorithm presented in this paper is experimentation on ORL library,MIT library and Yale library.The experimental results show that the average retrieval rate of the algorithm on 3 standard image library is 98.75%,96% and 97.3% respectively,and the maximum difference is 2.75%.It shows that the algorithm has good adaptability.