Research On Unsupervised Image Retrieval Based On Sparse Graph Structure

With the increasing of image data,image approximation search,which is also known as image retrieval,is playing an increasingly important role.In the past few years,supervised image retrieval has achieved satisfactory results.However,due to the scarcity of labeled data sets and the high cost of manual tagging,unsupervised retrieval technique has become a major concern.To save storage space,this paper studies the application of sparse graph structure in unsupervised image retrieval,aiming to improve the accuracy of image approximation search while saving memory.At present,this technology has the following three main problems to be solved:(1)The neighborhood information exploration in the case of lacking label information.So far,the performance of unsupervised image retrieval algorithms are still not as good as that of supervised ones overall.And the main reason is that the lack of label information leads to incomplete original neighborhood information,and thus leading to the degradation of the overall retrieval performance of unsupervised algorithms.Therefore,appropriate models are needed to explore and store approximation information in unsupervised algorithms.(2)The neighborhood information inconsistency in cross-modal retrieval tasks when the above problem solved.Due to the inconsistency of each modal data structure and the lack of label information,the neighborhood information of each modal is not the same exactly.Therefore,unified semantic neighborhood relationship must be achieved to guide the training process of cross-modal retrieval algorithm.(3)The efficient use of global neighborhood information in the training process of deep models when the above two problems solved.Since the achieved neighborhood information above is global and the optimization process of deep model is usually in mini-batch,the global information in a batch is destroyed.Therefore,the global neighborhood information should be made full use of to reduce the performance degradation in deep models.Based on the above three problems,the main contents and contributions of this paper are listed as following:(1)To solve the first problem,this paper proposes a sparse graph based self-supervised hashing algorithm.This algorithm uses sparse graph structure to store the explored neighborhood information and leverages it during training.In order to prove the rationality of the structure,the sparse graph in this algorithm is constructed by common k NN algorithm,and the same value of k is used for all experimental data sets,and the sparse graph is used in a linear algorithm.In addition,the algorithm introduces reconstruction loss and quantization loss to improve retrieval performance,and the experimental results on multiple data sets demonstrate the effectiveness of the algorithm.(2)According to the second problem,this paper puts forward a semantic rebasing based cross modal retrieval algorithm,and the algorithm uses k NN algorithm to build a sparse geometric neighborhood graph in each modal.To get unified semantic neighborhood information which can be shared by different modalities,a semantic sparse graph is introduced to provide weights between 0 and 1 for the neighborhood connection of each modal,representing the probability of the neighborhood relationship.As a result,the semantic probability sparse graph and the neighborhood geometric sparse graph of each modality constitute the neighborhood graph structures of proposed cross-modal retrieval algorithm.In addition to quantification and reconstruction loss,the algorithm also adds constraints on the weights of semantic probability sparse graph to achieve expected weight values.(3)As for the third problem,this paper proposes a sparse graph based deep hashing algorithm.This algorithm still uses the neighborhood sparse graph structure introduced by the above two algorithms to guide the training of the deep neural network.And this model makes changes in the traditional training scheme,so that weights of the neural network can be updated effectively by leveraging the neighborhood relationship in sparse graph during training process.For simplicity but without loss of generality,the algorithm is implemented on single modal data sets.Experimental results on multiple data sets show that this method can effectively reduce the performance degradation caused by lack of global information.Besides,due to the powerful fitting ability,the proposed model can achieve good performance in high relevance retrieval.