Design And Realization Of Art Image Retrieval System Based On Convolution Neural Network

With the development of deep learning,the research field of artificial intelligence moves to a deeper and more practical level,among which convolution neural network has very important research value in image feature expression.In this thesis,according to the national science and technology support project "National Arts and Crafts key supporting Technology Research and Application demonstration",the project resource library contains a large number of art images.A practical art image retrieval system is researched,designed and implemented.On the one hand,comparing with the traditional images,these art images have many kinds,various patterns,and high similarity.They are complexity and difficult to design and implement.On the other hand,as an important function of the project,image retrieval system has a strong practical value and research significance.In this thesis,the research and design of art image retrieval system is based on convolution neural network.The research work has been carried out in the following two aspects.One is the use of convolution neural network to extract the features of the art images.The scale of the art image data set in the database is relatively small.For the deep learning research field,the data set can only be considered as a small scale data set,and the hardware environment of the experiment is also relatively common.Considering the actual situation,the convolution neural network model is trained,and the better convolution neural network structure,which has been called ArtNet,is selected through experimental analysis.The convolution neural network structure of ArtNet consists of a seven-layer network structure.It is consisting of two convolution layers,two pool layers and one full connection layer.The experimental results show that the accuracy of ArtNet on the verification set can reach more than 94%.The other aspect is the design and implementation of a practical art image retrieval system based on ArtNet,which is an important function of the project construction and is effectively integrated with other systems of the project.The system consists of index creation and retrieval.The process of index creation is to extract the feature of the original image data set by using ArtNet.Each image gets a feature vector,which is used as the feature database of the original image data set,so these data constitute an index set.The retrieval process is to extract the feature vectors of the query images by using Art Net,then,we compare them with the feature vectors in the index set one by one,and calculate the similarity and then return the most similar images.The whole system is based on B/S structure,and the framework and language are Caffe and Python.The experimental results show that the accuracy of the retrieval system is above 70%.