Semantic Discrimination And Similarity Preservation Hashing:Research And Application

Computation speed of more than a billion images in seconds have resulted in the generation of binary code that has attracted great attention in recent years.The most popular application for generating binary code is image retrieval,which aims to re-turn semantically related images from a database via Hamming distance.As a kind of popular method for generating binary code,the hashing methods can be rough-ly divided into two categories:data-independent hashing(DIH)and learning-based hashing(LH).DIH usually generates hash functions by random projection,ensuring that neighbor samples can have similar binary codes with large probability.Although DIH is theoretically ensured to have high collision probability for similar data items,in practice it usually needs long hash bits and multiple hash tables to achieve both promising precision and recall.LH is known to perform much better than DIH with compact binary code.The ability of LH is to learn binary code by incorporating the structural information and distribution from training data.In visual tracking,appearance model is a key part of tracking algorithms.Template match is commonly used to calculate the confidence of candidate samples.However,this procedure is always time consuming,which leads the low speed of trackers.We observe that hashing techniques can be used to represent object by compact binary code which is efficient for processing.Thus,a multi-feature hashing tracking algorithm is proposed.In our tracker,the model fuses different.features to learn the hash functions,samples and templates are hashed to binary matrices,and the hamming distance is used to measure confidence of candidate samples.In addition,the designed incremental learning model is applied to update hash functions for both adapting situation change and saving training time.Experiments on our tracker and other state-of-the-art trackers demonstrate that the proposed algorithm is more robust in dealing with various types of scenarios.In the field of nearest neighbors search based image retrieval,most learning-based hashing algorithms leverage sample-to-sample similarities,such as neighborhood struc-ture,to generate binary codes,which achieve promising results for image retrieval.However,it is non-trivial to design sample-to-sample similarity to distinguish samples of similar classes.To address this issue,a class-level encoding method,which eencodes the class-to-class relationship,is used to take the semantic information of classes into consideration.Based on these two encodings,a novel framework,Error Correcting Input and Output Coding which does class-level and instance-level encoding under a unified mapping space,is proposed.The proposed mocdel contains two major compo-nents,which are distribution preservation and semantic encoding preservation.With these two components,our model maps the input feature of samples and the output code of classes into a unified space to encode the intrinsic structure of data and se-mantic information of classes simultaneously.Under this framework,Error Correcting Input and Output Hashing(EC-IOH)is presented via approximating the mapping s-pace with the Hamming space.Extensive experiments are conducted to evaluate the retrieval performance,and EC-IOH exhibits superior and competitive performances comparing with popular supervised and unsupervised hashing methods.Nearest neighbors search based image retrieval only can deal with the retrieval task with single semantic level.In practical applications,an image may belong to several categories at different levels.For example,an image of a boxer has a fine label boxer and a coarse label dog.Such hierarchical semantic information leads to our study of the fine to coarse retrieval(F2CR)problem,which aims to first retrieve samples with the same fine label for a given query,and then to return samples with the same coarse label.To evaluate the retrieval accuracy of F2CR.,a new metric called FC score,which takes into consideration both fine and coarse labels,is defined.Current hashing methods are not designed to efficiently handle the retrieval problem in this scenario.To solve F2CR,we propose a simple yet effective hashing method,dubbed fine to coarse hashing(FCH),in which FCH generates and modifies the output code of classes to encode the hierarchical semantic information.Extensive experiments show the effectiveness of FC score and demonstrates that our hashing approaches achieve very promising accuracy on the F2CR problem.With the continuous growth of data coming from the Internet,the number of images increases fastly.The coming samples will cause the changes of the original data distribution,which may lead the failure of initially trained hash functions.To alleviate this issue,online hashing was proposed.However,updating binary codes becomes more and more time-consuming.To deal with this procedure efficiently,a novel updating technique for hashing methods,namely Hash Selection Learning(HSL)is proposed.The motivation of HSL is to select representative hash functions from all original hash functions.Through HSL,the speed of updating binary codes for all samples is aimed to improve.A method for HSL,which based on greedy selection strategy and the Distribution Preserving Hash Selection learning(DHSL)algorithm,is presented by designing a novel loss function.The experimental results demonstrate that the HSL is effective to improve the speed of online updating and the chosen hash functions can furthest preserve the discriminability of original hash functions.Finally,the summarizations of relevance,applicability,advantages and disadvan-tages of the proposed problems and methods are presented.Based on these summa-rizations,several possible future work is introduced in the final chapter.