Hashing Learning For Large-Scale Image Retrieval

Hashing learning can significantly improve retrieval speed,reduce memory space and communication cost in large-scale image retrieval,which make it become a research hotspot in image retrieval.The core idea of hashing learning in image retrieval is to map each image in the image database into a compact binary code sequence by learning an explicit(or implicit)hash map,so that the similarity between the hash codes of similar images is higher(ie,the number of different binary bits is small).We only need to calculate the Hamming distance between the hash codes to quickly compare the similarity of images.Meanwhile,it saves memory space and improves communication efficiency.It has been a difficult task that searching quickly one or more images most similar to query images from a massive high-dimensional image database in large-scale image retrieval.Therefore,it is urgent to explore new technologies to achieve large scale image retrieval with lower storage cost,higher retrieval efficiency and precision rate.We present image retrieval algorithms base on multi-feature and multi-kernel hashing and multi-feature spectral hashing by introducing multi-feature fusion and multi-kernel learning into hashing learning,and verify the retrieval performance in both theory and experiment in this thesis.The specific work is as follows:1.This thesis firstly compares the hashing learning methods in image retrieval and propose a large scale image retrieval framework base on hashing learning.The motivation,model and algorithm,similarities and differences,advantages and disadvantages of the methods are analyzed from the perspective of kernel hashing,deep hashing,manifold hashing and sparse hashing.Then,the identical class of methods are evaluated,which are classified according to their generality.Finally,the procedure of hashing learning for image retrieval is generalized.2.This thesis proposes a large-scale image retrieval algorithm base on multi-feature and multi-kernel hashing learning(MFMKH).Kernel hashing choose one or two features corresponding to only Gaussian kernel,ignoring the advantages of multi-feature fusion and multi-kernel learning.Further,they cannot fully utilize the learning ability of different kernel functions and different information contained in different features.This thesis employs feature fusion to not only address the problem of insufficient information with single feature but also find the intrinsic relation between different features.Multi-kernel learning can not only solve the problem of “dimension disaster” and the problem of linear inseparability with the kernel function,but also can overcome the problem of exorbitant dimension in high multi-feature fusion.3.This thesis proposes a large-scale image retrieval algorithm base on multi-feature spectral hashing learning(MFSH).Due to the fact that supervised hashing suffers from obtaining labels in high cost,spectral hashing uses single feature without sufficient information and multi-kernel hashing has too long training time,this thesis resolves the problem of lack of labels by constructing the affinity matrix,solves the problems of single features without sufficient information by the fusing local features and global features,and solves the problem of long training time by using the Laplacian eigenvalue decomposition method.