Image Retrieval Based On Deep Learning

With the arrival of big data and artificial intelligence,image data begins to grow explosively,and efficient retrieval of massive data has become an urgent need of modern society.Content-based image retrieval can play a key role.Although using the deep features extracted from the convolutional neural network directly can abstractly describe the content of the image,it has the problem that the search performance is not high and the local details are not prominent enough.Considering of the above mentioned problems,this thesis focuses on how to use deep learning algorithm to achieve efficient image search and detail representation of interested contents,and studies related content from two aspects of similar image retrieval and instance-level image retrieval.The specific research contents are as follows:Firstly,to improve the search performance optimization problem of similar image retrieval methods,deep hash algorithm is designed from three aspects: the improvement of softsign activation function,and the design of loss function for image pair and learning strategy,so that similar images can be efficiently encoded into similar hash binary codes.The experimental results show that,compared with other methods,this method can achieve higher performance for similar image search.And then,at the aspect of the instance-level image retrieval method,to solve the problem of detailed description of specific objects,structures and scenarios,this thesis gets the saliency map with semantic information by using class activation map.Based on spatial and channel weights of saliency map,deep features that can express details of interest are extracted to make region of interest more salient.The experimental results show that the method can depict the local details of the image more prominently and have a higher retrieval accuracy rate.Finally,based on the above research,we design and complete an image retrieval software which can complete image search faster and better,and meet the need of image retrieval visualization.In this thesis,deep learning algorithm is used to obtain image representation,to achieve efficient search of similar images,and to extract the detailed features of local interest in the image.It provides image retrieval methods which have much higher retrieval accuracy rate and faster retrieval of massive image data.