Research On Image Retrieval Algorithm Based On Full-Connected Weight Vector And MCR-MAC Feature

With the rapid development of Internet media era,image data as the mainstream carrier of media information,its data scale grows rapidly with the rapid upgrade of intelligent devices.With the rapid development of deep learning technology,deep convolution features with high-level semantic information have gradually replaced traditional manual features and become a research hotspot in the field of image retrieval.Although the deep convolution feature contains more abundant feature information,it still needs to be compressed into low-dimensional vectors by the feature compression method before it can be used in image retrieval.Therefore,it is an important breakthrough in the field of image retrieval to compress the depth feature and obtain effective image representation at the same time.In order to realize the compression of deep convolutional features and obtain robust low-dimensional features,this paper draws on the pooling idea of R-MAC(Regional Maximum Activation of Convolutions)sliding Windows,and enshrine regional weights by improving the pooling response of feature graphs.The characteristics of MCR-MAC(Maximum Correction and Regionalweighted Mixing-pooled Activation of Convolution)are proposed.On this basis,the full connection layer is introduced to give weight to the channels with MCR-MAC features,which greatly improves the accuracy of image retrieval.The main research contents of this thesis are as follows:(1)Referring to the idea of R-MAC sliding window,the maximum response value of the feature graph was modified based on the sampling frequency of two-dimensional prefix and matrix statistics,and combined with the average response and maximum response of the sliding window region to construct the regional weight,a convolutional feature post-processing method of maximum correction and region-weighted hybrid pooling convolution activation(MCR-MAC)was proposed.The parameters of the model were fine-tuned combined with the improved terre loss function.Compared with the original R-MAC features,the m AP of the Oxford5 k data set was increased by 13.7%and 11.4%,respectively,and the m AP of the Paris6 k data set was increased by 7.6% and1.2%,respectively,under the input of the two resolutions.(2)On the basis of the MCR-MAC feature,the output vector of the full connection layer is introduced,and each channel of the MCR-MAC feature graph is Weighted by the inner product of the two,and a weighted weighted MCR-MAC feature is proposed.Adaptive pooling layer is introduced to solve the problem of mismatch of full connection parameters under different resolutions.The parameters of the convolution layer are frozen,and the improved ternary loss function is used to train the parameters of the full connection layer of the model.Combined with extended query,compared with the MCR-MAC feature,the improved method increased the m AP of the Oxford5 k dataset by 16.5% and 9.0% for the two resolution inputs,and increased the m AP of the Paris6 k dataset by 8.2% and 5.8% for the two resolution inputs,respectively.