Research On Unsupervised Hashing Methods Based On Similarity Preserving

With the rapid development and widespread application of information technology, image data has been widely disseminated and it continues to grow every day. The upgrading of the mode of image data acquired also makes mages become broad, diverse, complex and high-dimensional.. It is undoubtedly that these features lead to enormous challenges for image retrieval and storage. How to extract useful image information in these massive, high-dimensional, complex image data become a hot topic. What's more, how to extract useful image quickly and efficiently become the core of research questions. To solve the above problems, the hashing algorithm has drawn wide attention as an efficient and fast image retrieval technology in recent years. Hashing can transform an image into compact binary coding sequence so that the high-dimensional data in original space embedded into low-dimension Hamming space,which can greatly improve the retrieval speed and save storage space.Based on depth study of the existing hash algorithms, in order to maintain similarity structure among data points in the process of mapping and quantitative effectively. First, inspired by the signal to noise ratio(SNR) function, this paper introduce an effective feature extraction framework. We use conditional second moment matrices to construct signal and noise vectors for every cluster data points, then to get discriminative information of original data through solving the ratio of signal to noise using generalized eigen-decomposition.Second, a linear embedding hashing is proposed, this method is intended to improve the correlation between the image and it's coding so that to achieve the better precision of image retrieval. This method introduces similarity predicted function to represent correlation between high dimensional data and its codes, then the optimal embedded matrix can be obtained by minimizing margin loss function, which can maintain affinity relationships between high dimensional data and its codes, thus achieves better binary encoding to database.Finally, compressed hashing for neighborhood structure preserving is proposed in the paper. First, It introduces stochastic neighbor embedding algorithm to learn a new set of synthetic reference vectors represent database in the original space, which can avoid the infeasibility when finding data point's nearest neighbors at the large scale database. In addition, we exploit the manifold learning to generate appropriate hash function, which preserve the intrinsic structure of high-dimensional data points in the low-dimensional manifold space and solve the drawbacks of partial information loss. And the hashing method proposed in this paper obtains very high precision and recall.Experimental results show that the linear embedded hash algorithm improved the correlation between the image data and codes, and realized the high precision of performance in image retrieval. What's more, compressed hashing for neighborhood structure preserving effective preserve the intrinsic structure of high-dimensional data points in the low-dimensional manifold space so as to the precision of image retrieval is better than other hash algorithms.