Research On Cross-modal Hash Retrieval Technology And System Design

In recent years,with the rapid growth of multimedia data,cross-modal retrieval plays a significant role in massive data processing.Therefore,people pay more and more attention to the development of cross-modal retrieval technology.The main task of cross-modal retrieval is to obtain semantically related data in one modality for a query in another modality.For example,with a textual query,the user can get not only documents,but also some semantic relevant pictures or videos.It demands efficient solutions in large-scale image and text data similarity retrieval applications.It can effectively improve retrieval speed and reduce storage space to convert data into compact binary code.Due to the computational efficiency and low storage consumption of hash,the approximate neighbor search based on hash technology is more and more widely used in multimedia data retrieval processing.Most of the existing hash-based cross-modal retrieval techniques are supervised methods,utilizing the tags carried by the data to train the model.In practice,the labels of the data are manually labeled,so the supervised method requires a large number of tags,which is not suitable for practical large-scale retrieval applications.The unsupervised method can effectively solve this challenge.How to make similar data points still similar while mapping to the corresponding hash code is the main challenge faced by hash-based approximate neighbor retrieval technology.Based on hash technology for cross-modal retrieval should preserve the inter-media and intra-media semantic similarity effectively.The existing hash-based cross-modal retrieval technology generally relaxes the hash code to a continuous value in the optimization process,and finally obtains a hash code by setting a threshold.However,this method may cause a lot of useful information to be lost such that the quality of the learned hash code is degraded.In order to solve the above problems,this paper proposes an cross-modal hash method called hypergraph-based discrete hash learning(BGDH),which belongs to unsupervised learning method.The novel method utilizes a hypergraph to capture the high-order relations among instances which can preserve intra-media semantic similarity for cross-modal retrieval.We propose an efficient discrete hash optimization technique to directly solve the hash codes and reduce the quantization information loss.In addition,extensive experiments on three benchmark datasets demonstrate the state-of-the-art performance of our method and validate the desirable advantage on boosting cross-modal retrieval performance.Finally,we design a cross-modal retrieval system,which has two main retrieval function.The system consists of input module,cross-modal retrieval module,file management module,output module,user management module and page display module.The core module of the system is the cross-modal retrieval module.The input module of the system performs featureextraction on the data that uploaded by the user and then calls the cross-modal retrieval module.The core technology in the cross-modal retrieval module includes feature transformation and similarity search.The cross-modal retrieval module can realize the cross-modal retrieval function.The system chooses Mysql as the system data management tool in the system database design to ensure the reliability of the system data.The system runs in C/S mode.