Image Semantic Consistent Hashing Representation For Image Retrieval

As a classical approximate nearest neighbor algorithm,hash algorithm has attracted the attention of a large number of researchers in the era of big data by virtue of its efficiency advantages.Recently,the existing hash algorithms use the expression ability of deep neural network to construct accurate hash mapping through a large number of supervised data training models.However,the strong dependence on a large number of supervised data limits the further implementation of deep hash algorithm.Therefore,in the case of insufficient label data,this paper studies the hash coding representation of semantic consistency mapping.The research of this paper is of great significance for the practical application of hash algorithm.After analyzing the existing hash coding algorithms,it is found that there are three main problems: First,image semantic consistency is difficult to map to coding space.Second,the accuracy of mapping image semantic consistency to coding space is insufficient.Complex background,sample imbalance and other interference factors will lead to mapping errors.Thirdly,the accuracy of mapping image semantic consistency to coding space is difficult to maintain in a dynamic environment.This paper gives corresponding solutions to the above problems in the analysis.Specifically,the research contents and main contributions of this paper include the following four aspects:Firstly,in order to map the image semantic consistency to the coding space,this paper proposes a balanced discrete hash algorithm based on the principle of semantic consistency mapping.The principle of semantic consistency mapping requires mapping to ensure that the semantic relative relationship of samples remains unchanged.Balanced discrete hash constructs hash mapping through the semantic relationship between coding layer and feature layer.Specifically,the proposed algorithm uses graph structure to establish the semantic relationship model of feature layer.The degree of separation between features is introduced into the definition of edges,and the coding is used as the node of the graph.Then,this paper proposes the consistency loss of orthogonal timing.The loss term not only reduces the redundancy of coding,but also further modifies the hash map through self-monitoring.Finally,the discrete cyclic coordinate descent method is used to solve the target problem,which avoids the semantic loss in the quantization process.Experiments show that the proposed algorithm establishes a mapping between category semantics and coding without relying on label information.Secondly,in order to map more fine-grained image semantic consistency to the coding space,combined with the principle of semantic discriminant mapping,this paper proposes a local discriminant hash algorithm.The principle of semantic discriminant mapping requires mapping to separate different sample points as much as possible and compact the same ones as much as possible.Local discriminant hash uses a small number of label samples and requires hash mapping to focus on the local semantic region of the sample,so as to increase the degree of intra class aggregation of coding,so as to construct hash mapping.The proposed algorithm includes semantic mining model and hash model.Among them,hash model is used to supervise discriminant mining.Specifically,this paper proposes a discriminant loss based on quads,which requires the samples after discriminant processing to be mapped to hash codes with higher discrimination.Experimental results show that the proposed algorithm can establish a mapping between the semantics and coding of fine-grained categories.Thirdly,in order to improve the accuracy of image semantic consistency mapping to coding space,this paper proposes an anti generation hash algorithm based on semantic mapping self-step expansion.The proposed algorithm imitates the human learning process and gradually improves the difficult degree of hard samples according to the accuracy of hashing.Difficult samples refer to samples that are difficult to encode,which results from the loss of semantic region and sample imbalance,that is,some samples are ignored in the training process due to insufficient quantity.To this end,self-step confrontation generates hash,joint learning difficulty sample generation model and hash model.Generate appropriate difficult samples according to the requirements of hash model.Firstly,an efficient generation model is designed to generate difficult samples from occlusion and deformation respectively.Then,this paper proposes a generation strategy based on selflearning to control the generation of difficult samples from easy to difficult.In addition,this paper proposes the loss of semantic consistency to mine the semantic information of unsupervised samples and improve the accuracy of hash mapping.Experiments show that the proposed algorithm can establish more accurate and robust hash mapping.Finally,in order to maintain the accuracy of image semantic consistency mapping to coding space in dynamic environment,a central adaptive hash algorithm for real-time adjustment of semantic consistency mapping is proposed.The algorithm process includes two stages: building the category center adaptively for the new category,and then fitting the sample to the specified category center by the hash model.Firstly,this paper uses the standard orthogonality of unitary matrix to construct a category center with large enough discrimination.Secondly,in order to retain the semantic memory of old samples,this paper constructs a distillation learning framework based on global semantic knowledge.And through the strategy of asynchronous update to alleviate the problem of catastrophic forgetting.Finally,for the new input samples,this paper proposes a bit level attention loss,which takes the probability density as the weight of the samples to make the model focus on the samples with poor fitting in real time.Experiments show that the proposed algorithm can adjust the mapping in real time to maintain accuracy in a dynamic online environment.Through the above research,this paper makes an in-depth exploration on the hash algorithm of semantic consistency mapping in image retrieval,provides practical and effective solutions to the key problems,and makes targeted improvement and innovation on the online application of the algorithm.The algorithm proposed in this paper has important research significance for the practical application of hash algorithm.