Image Retrieval With Deep Learning And Hashing

Content-based image retrieval has received widespread attention in the visual field.Retrieval is mainly divided into two steps.First,the features of the image are extracted,then look up the image in the database which feature approximates the feature of the query image.The most important step is the features extraction of the image.With the incresaing amount of image data,we hope that the image features used for retrieval can contain sufficient information,that is,features that could have semantic information.Secondly,that the features dimensions could be low,which can reduce the storage cost of the features.And finally,we hope that the features are sufficient simple to speed up retrieval.In recent years,deep convolutional networks based on learning have made great achievements in the field of computer vision.The features extracted by the convolutional neural network can well express the content of a picture,and the convolutional network can also extract low-level images local fetures and high-leval semantic features.More and more image retrieval algorithms based on deep learning are being proposed.This article is based on thinking about the characteristics of image retrieval,so the follwing parts of the research work are carried out:1.To study the image retrieval algorithm based on deep convolutional neural network feature fusion,which aiming at riching features content.Firstly,the features of each layer of the convolutional layer and the features extracted by the different kernels are visualized.Then,analyse the different characteristics of these features.The two methods of feature fusion are studied:(1)Feature fusion based on features of different layers.(2)Feature fusion based on features extraction by differnent scale kernels.2.To study the deep random VLAD-based hash retrieval algorithm(RV-SSDH),which aiming at reducing the features dimension and complexity.Firstly,the classic convolutional neural network is used to extract the convolutional features of theimage,and then the features are reduced dimension by random-VLAD module.The reduced-dimensional features are hash-encoded,and finally a low-dimensionalhash code is generated.While reducing the features dimensions,the algorithm also useds a hash algorithm to reduce the complexity of the features,and the distance between features can be calclated by the Hamming distance,while greatly speeds up the algorithms' s retrieval speed.Compared with other algorithms,it shows the advantages of RV-SSDH.