Research On Hash Algorithm In Image Retrieval Tasks

Benefiting from the development of big data and cloud computing techniques,massive data are gradually flooding the Internet and user terminals,forming multimedia data resources with large volume,complex data structure and high processing difficulty.Among them,image data with its representative semantic feature content,data internal correlation,extensive application scenarios,and diversified retrieval and recognition requirements,has become a representative category in multimedia data retrieval and recognition tasks.When dealing with the problem of multimedia information retrieval,high time and space complexity because of the high data dimension often occurs in the retrieval processing.At this time,using linear search method to retrieve and identify data will be laborious.Approximate Nearest Neighbor Search,ANNS,are used to significantly reduce the processing complexity without significantly reducing the retrieval performance.Among ANNS techniques,retrieval hash algorithm has attracted more and more attention because of its advantages of efficiency,accuracy and robustness.Retrieval hash algorithm uses the collision concept of hash mapping to make relevant data have the same or similar hash codes,which enable fast and accurate ANNS via matching hash codes.CBIR systems based on retrieval hash algorithm tend to have high retrieval speed and accuracy compared with traditional retrieval algorithms.Relevance feedback and deep learning methods can further improve the retrieval performance of retrieval hash algorithm in different application scenarios.Focusing on the application of hash algorithms in image retrieval tasks,in this dissertation we proposed a series of application,improvement and expansion modes of hash algorithm in the field of image retrieval and recognition for different types of image retrieval and recognition tasks.Our work focuses on these problems and presents the following contributions:1.Aiming at the characteristics of distributed image retrieval tasks,based on retrieval hashing algorithm and distributed hash table structures,a multi-feature relevance feedback based image retrieval algorithm is presented for distributed retrieval system.The retrieval system uses multi-texton histogram and SURF feature to represent image contents,uses multiple hash buckets to map image data,and stores the mapping result into DHT structures to implement fast nearest neighbor search in distributed scenarios.By using relevance feedback strategy to distill retrieval results generated using different image features,the retrieval algorithm enables retrieval system provide results representing user intentions.The retrieval algorithms are designed and applied in a distributed image retrieval system.Compare to existing algorithms,the multi-feature relevance feedback based image retrieval algorithm can acquires better retrieval performance while maintaining the load balance performance of DSH structures.2.Aiming at the problems in designing loss functions for fully end-to-end deep hashing algorithms,an integrated regularization based deep hashing algorithm DSHIR(Deep Supervised Hashing Network with Integrated Regularization)is proposed.The integrated regularization deep supervised hashing algorithm mainly focuses on two different learning problems which are deep feature learning and hash code generating.By systematically studying the similarities and differences of these two problems,DHSIR optimized the similarity preserving subloss function in deep hashing algorithms,raised novel idea of zero division regularization loss,and discussed the actual function of binary output restriction loss in existing deep hashing algorithms.The designs in DSHIR algorithm boosted the performance of fully end-to-end deep hashing algorithm,and several expansive conclusions are obtained.3.Aiming at parameter preloaded deep hashing algorithms for multi-label image retrieval scenarios,a multi-label distribution optimization algorithm,DSH-COS(Deep Supervised Hashing with label Co-Occurrence Similarity)is proposed.Hashing algorithm provides fault tolerance ability for ANNS systems,which is revealed differently in single-label and multi-label scenarios.DSH-COS algorithm utilizing this difference,modeled the relationship between multi-label images based on label relation,and by designing individual similarity computing module outside the deep structures,established a highly effective and flixible deep hashing algorithm for ANNS task.Evaluation on standard datasets shows DSH-COS algorithm yields better retrieval performances compared to existing algorithms,and the image similarity modeling method used in DSH-COS algorithm is advantageous.4.Aiming at the problems in large scale ANNS tasks with Product Quantization encoding,the PQMT(Product Quantizaiton based Multi Table retrieval)algorithm based on multi-table result fusion design is proposed.By studying the performance bottlenecks in state-of-the-art techniques,PQMT raises novel multi-table result fusion algorithm with more efficiency and robustness,enable the retrieval algorithm provide faster query processing ability while maintaining the same recall rate.Compared to existing retrieval algorithms,PQMT algorithm yields significantly faster retrieval speed while maintaining the recall rate.