Research On Multi-label Image Retrieval Method Combined With Convolution Neural Network

In recent years,the research of image retrieval is maturing.However,there are still many urgent problems to be solved of image retrieval,such as eliminating the semantic gap,reducing the dimension of image features,and improving the speed of image retrieval.At first,researchers applied hashing methods to image retrieval to effectively alleviate the problem of image feature storage and computation,but it also led to the difference between hashing code and image features.In recent years,deep learning is integrated into image retrieval by researchers,which makes it possible to extract the deeper features of the image.The image retrieval model which is based on convolutional neural network and hashing method is the trend in development of image retrieval,but this model is not mature.Moreover,the convolutional neural network is in the development stage.It needs to overcome some problems,such as the construction of models,the selection and optimization of model parameters,and semantic gap.In this thesis,a suitable image retrieval framework is designed based on the above problems.The main work of the thesis is as follows:(1)The working principle and optimization method of convolution neural network are studied and analysed.At the same time,a variety of algorithms about hash function are researched.(2)In this thesis,the traditional image retrieval model VGG which is used for feature extraction is improved.This research use any two image labels to generate an image pairs label as an expected value,and converts the feature value of this two images into an image pairs label as actual value.The training condition of the network is changed from using the feature and label of a single image to using the actual and expected values of an image pair.(3)In this thesis,the traditional feature extraction network is improved.Based on the VGG model,the network architecture is improved.The traditional volume layer and the full connection layer are replaced by the multilayer perceptron and the global average pool layer,and the image feature is extracted through this network.The improved feature extraction network parameters are greatly reduced,the network structure is simplified,and the retrieval speed is improved.(4)In this thesis,the network learning algorithm is improved.The network learning algorithm consists of three parts: hash function,loss function and random descent gradient algorithm.This research improved the loss function in the learning algorithm.The loss function in this research includes not only the loss term of the expected value and the actual value,but also the difference between the feature veators and the hashing code,and the difference between the feature value and the mean value of the image according to the stationarity of image.This learning algorithm not only reduces the difference between the target value and the expected value,but also alleviates the semantic gap.No matter the image is hashed or eigenvalue,it can ensure the smoothness of the image.The above research points are integrated into a multi-label image retrieval framework.Using CIFAR-10 and NUS-WIDE sets to test the framework,compared with other related methods,the speed and MAP of framework are improved.