| Big data has become one of the country’s important strategic resources.The storage,processing,retrieval and analysis of big data are important links in big data processing.Approximate nearest neighbor retrieval is an important way to solve big data retrieval problems,and hash learning is a representative method to solve big data retrieval and analysis-based problems,which has received extensive attention from researchers in recent years.Using hash learning for big data retrieval can greatly reduce data storage and computing overhead,and has unique advantages in big data retrieval and analysis applications.Although the research on hash learning has made fruitful progress in recent years,there are still some deficiencies and problems worthy of further research:1)Most of the existing hash learning methods learn the hash code from the semantic information of the whole image,ignoring the influence of different regions of the image on the hash code;2)Most of the existing studies use a predefined similarity matrix to learn hash codes,but this similarity matrix cannot accurately reflect the difference of sample similarity;3)Existing work ignores the mining of difficult samples in complex scenes,and does not pay attention to the difference in learning difficulty of different samples;4)Most of the common hash code quantization methods assign a hash code with the same length to each dimension,and cannot distinguish the difference in the amount of information of the features of each dimension.In order to solve the above problems,this thesis conducts a more in-depth study on the hash learning method using the idea of weight information mining.Five hash learning methods are proposed from four aspects:sample space importance weight information mining,sample similarity matrix weight information mining,difficult sample weight information mining and different dimension weighted assignment of hash codes.The details are as follows:(1)Aiming at the problem that the existing hash learning methods ignore the importance difference of different regions of the image,this thesis proposed the deep hashing with weighted spatial importance.The method first uses a spatial attention model to learn the importance of different spatial regions in the original image,then assigns different weights to these spatial regions according to their importance,and learns hash codes based on the weighted spatial information,which improves the depth of hashing.performance of the method.(2)Aiming at the problem that special supervised information is difficult to obtain in the zero-shot hash learning method,thesis dissertation proposed the zero-shot hashing via asymmetric ratio similarity matrix.This method uses an asymmetric ratio similarity matrix(that is,the positive weights of similar sample pairs are not equal to the negative weights of dissimilar sample pairs)in the sample similarity calculation to explore the influence of supervised information on hash learning.And only need to calculate the semantic similarity between seen classes for hash learning without special supervised information.To a certain extent,the dependence of zero-shot hash learning on special supervision information is alleviated.(3)Aiming at the problem of insufficient utilization of supervised information by existing hash learning methods,this thesis further proposed the supervised adaptive similarity matrix hashing.This method can not only learn the similarity matrix and hash code adaptively,but also can extract the label correlation by maintaining the consistency between the feature and label space,thereby effectively improving the retrieval performance of hash learning.(4)In view of the insufficient mining and use of difficult samples by existing hash learning methods,this thesis proposed the deep circle hashing via adaptive similarity matrix.When dealing with retrieval problems in complex scenes,more learning weights can be assigned to difficult samples.And in the loss function design,the characteristics of adaptive similarity matrix,hash similarity,interval weighting and hash retrieval task are introduced,which improves the retrieval accuracy of hash learning method in complex scenarios.(5)Aiming at the problem that the existing hash quantization methods cannot adaptively allocate the length of the hash code,this thesis proposed the variable-length quantization strategy for hashing.The method optimally allocates different numbers of bits to each dimension after image projection to obtain minimum dispersion,thus effectively improving the performance of hash quantization steps.The proposed five methods introduce the idea of weight information mining to focus on distinguishing important information.The experimental results show that making full use of the weight information mining idea is an effective means to improve the performance of hash learning methods. |
