Research On Image Retrieval Method Combined With Hashing And Deep Learning

With the rapid development of multimedia technology,the traditional information retrieval technology has become more and more difficult to meet the people's demand for information retrieval.As the result,the Content Based Image Retrieval(CBIR)system came into being.In the era of big data,how the content-based image retrieval run based on the large scale image data sets is the focus of research domain,and the key issues are higher precision and efficiency of retrieval system.The CBIR system needs to find ways to represent the image(Image Presentation),building indexes for image data.From the perspective of human recognition,the semantic information recognized by human is different from features extracted by machine,the ‘semantic gap' exists between machine recognition and human understanding.It makes a deviation between retrieval results and retrieval needs.With the develop Deep Learning research,Convolutional Neural Network(CNN)is applied to the field of image retrieval which has achieved good performance,but the information extracted from CNN have high dimension,which affect the efficiency of image retrieval.On the other hand,hash coding is the main method to realize fast retrieval in large scale image retrieval system.By binary-encoding of the image feature,it can improve the retrieval speed and reduce the memory consumption of data storage,helping to construct the efficient data structure.However,original information may lost by Hash Coding,as the result the performance of retrieval accuracy will decrease.It is necessary to solve the problem that the similarity changes after hash coding.In view of the above problems,this thesis realized an image retrieval method combined with Deep Learning and Hash Coding,and carried out the experiment and the analysis,the main work as follows:1.A deep learning method is studied.A Convolutional Neural Network model based on VGG-Net is constructed.The image database is applied to pre-training and tuning for the model.After pre-training,the model is used for image representation.2.The dimension reduction methods of Metric Learning and hash coding methods are studied.This thesis proposes an improved iterative quantization hash coding learning methods named SPCA-ITQ,which can reduce the information dimension and improve the performance of image retrieval.3.The image feature extraction module,hash coding module,image retrieval simulation module and evaluation module are designed and implemented,the method of this thesis is compared with various mainstream methods,and the results are analyzed.