Image Retrieval Problems In Non-stationary Environment

With the growth of imaging devices,the amount of images on the Internet grows explosively.Large scale image retrieval techniques have large research and practical values and are widely applied in fields of security,entertainments,and social applications.Current large scale image retrieval techniques are mostly based on hashing methods,which train a set of hash functions to project data from a high dimensional feature space to a low dimensional binary Hamming space.Compact hash codes are generated for each image.The similarity between images is finally evaluated based on Hamming distances between their hash codes.Various hashing methods have been proposed and achieve promising retrieval performance.However,most of existing hashing methods are proposed for stationary data environments,while data environments in real world are usually non-stationary with new images appearing over time.Moreover,the data distribution of class(es)may change over time,which is called concept drift phenomenon.The previously learned hash functions based on the old data environment cannot adapt to the new data environment well.Hash codes generated for new images cannot preserve similarities among images,which leads to a fading of retrieval performance.Therefore,this thesis focuses how to update hash functions in non-stationary environment with concept drift to keep both accuracies and efficiencies of hashing methods.Four dynamic hashing methods are proposed in this thesis to solve image retrieval problems in non-stationary data environment with concept drift.Major contributions of this thesis are summarized as follows:1.An Incremental Hashing method(ICH)is proposed using multiple hash tables.The similarity information in data chunks of different time steps is maintained in multiple hash tables.Weight is assigned to each hash table based on its similarity consistency and partition balancing of hash functions.The retrieval results of each hash table are merged using weights of hash tables to obtain the final retrieval results.To our best knowledge,the ICH is the first hashing method to solve the image retrieval problem in non-stationary data environment with concept drift.2.The Incremental hash-Bit Learning(IBL)is proposed for selection of hash functions in non-stationary environments.With the similarity information in new data chunk,new hash functions are trained and added to the hash function pool.The optimal combination of hash functions is selected based on the retrieval performance of each hash function to the current data environment.In addition to the similarity consistency and partition balancing of hash functions,the correlation between hash functions is also considered to avoid information redundancy.3.The Concept Preserving Hashing(CPH)is proposed to project new images with concept drift to original concept without updating hash codes of all images.The objective function in CPH consists of three parts: isomorphic similarity,partition balancing,and heterogeneous similarity fitness.New hash functions are learned by optimizing the objective function.To our best knowledge,the CPH is the first non-stationary hashing method which learns new hash functions without updating hash codes of all images.4.The Complementary Incremental Hashing method(CIHR)with query adaptive weighting is proposed in this thesis.The aforementioned ICH method trains hash tables based on data chunks of different time independently,without considering the correlation between hash tables.The CIHR trains new hash table based on the similarity information in new data and errors caused by previously learned hash tables,which makes the hash tables complementary.Moreover,in CIHR,query-adaptive weight is assigned to each hash function for a given query.Finally,weighted Hamming distance is used for similarity evaluation.Overall,the four proposed hashing methods in this thesis form a new research direction of non-stationary hashing in the image retrieval area.These researches are expected to inspire more interesting and useful follow-up researches and practical applications in this area.