Image Indexing Methods Based On Hashing

With the increasing popularity of smart devices, type and number of images showing explosive growth, the effective retrieval of images has become a hot topic.Effective image indexing is an important prerequisite for image retrieval,and the traditional image indexing algorithm is generally suitable for small and low-dimensional scenes, can't handle massive data. Therefore, for the massive and high-dimensional image data, the establishment of effective index and achieve accurate and fast retrieval task has become a research focus in computer vision. The image indexing algorithm based on hash has been widely concerned in recent years because of its high query speed and small storage. In this context, the main work of this thesis is to study hash algorithm and its application in large scale image indexing,including:(1) We propose an isotropic iterative quantization hash algorithm. The constraints imposed by the Iterative Quantization (ITQ) on the rotation matrix are too thin and easily lead to overfitting; Isotropic Hash (IsoHash) lack of update strategy of hash coding, which reduces the coding quality and so on. By using the iterative strategy, the coding matrix and the rotation matrix are alternately updated,and the isotropic constraints are added on the basis of orthogonal constraints to study the optimal rotation matrix, which makes the hash projection has a smaller quantization error. Compared with the mainstream algorithm, the results show that the algorithm has better precision and recall rate.(2) We propose a reverse spectral hashing. The optimization of the objective function of Spectral Hash (SH) can't guarantee that the neighboring samples in the original high dimensional space are still in close proximity to the low-dimensional Hamming space, and lag behind the mainstream Hash algorithm on various kinds of retrieval evaluation. In this thesis,the input and output positions are exchanged on the basis of SH, and the similarity definition is transformed from the original high dimensional data to the low dimensional hash code. The experimental results show that the proposed algorithm can ensure that Hamming distance is small after the adjacent sample projection, and obtain a better retrieval performance.(3) We design a web clothing search based on hash indexing. Users upload pictures,the server extracts features,codes,rearranges and a series of operations,and return the search results according to user needs. After a large number of tests,web clothing retrieval system has achieved satisfactory results.