Research On Image Retrieval Based On Hash Coding

In recent years,with the rapid development of mobile Internet and multimedia applications,image data has become an important part of the network data,however,high dimensional characteristics of the image data makes fast retrieval and storage of massive image data in network become a difficult problem which we are facing to.In order to solve this problem,the image hash technology was born and relied on its excellent data reduction ability,it has attracted much researchers' attention.Quality of hash code generated by image hash depends on the extracted image features.Because the traditional hash method which is based on artificial design features owns the limited capacity of feature extraction,the effect is not ideal in the practical application.With double-quick progress of deep learning,feature extraction method based on deep learning has gradually replaced the traditional method,a research hotspot of image hash technology is depth hash method that combines the deep learning with hash code.This method by virtue of its strong image feature and hash learning ability has achieved significant results in image retrieval field based on content.Although the research on depth hash method has achieved certain results,the more suitable network structure and optimization algorithms to this task has to be further explored.According to comprehensive analysis of the current situation of image hash technology at home and abroad,this paper mainly focus on research about depth hash method which is applied to image retrieval field.The main work of this paper is as follows:Firstly,the quality of the image is directly affected by the feature extraction.In order to get a higher quality of hash coding,this paper improves the CNN-F network to improve the ability of image feature extraction.The main improvement is as follows: improve the network characteristics of small objects extraction ability by improving the structure of CNN-F network;by introducing the temporal and spatial sampling layer in Pyramid,the feature extraction ability of the network for different size images is improved,and the network learning speed is accelerated.Secondly,we use the improved objective function to study the hash code,which reduces the error in the discrete optimization process and improves hash code generating quality.Then,this paper summarized and implemented the typical hashing method,the objective function using the improved CNN-F neural network and the improved depth to construct hash model.The hash function,the loss function and the network propagation algorithm are designed and implemented,the hash model is used to study the image features and hash codes to get better hash codes.In addition,this paper proposes an image retrieval algorithm based on the weight of the hash code,which can improve the accuracy of image retrieval by applying twice ranking retrieval on Hamming distance.Finally,the typical model and the deep hash method are implemented on the Cifar-10 and Nus-wide image data set.The results show that the retrieval efficiency of the proposed deep hash model is superior to the traditional one,and it also has some advantages over the recently proposed deep hashing method.