Large-Scale Visual Retrieval Based On Hash Learning

Nowadays, network has become one of the most important approaches to acquire information. Meanwhile, with the coming of big data era, multimedia data grows explosively. Since the 1990 s, people have developed a large number of work on the efficient storage and fast retrieval for multimedia data, and proposed successively text-based and content-based retrieval methods. Although these methods are easy to operate, due to the heterogeneous structure, diverse semantics, sophisticated content of the multimedia data, the methods are not suitable to such cases. Thus, how to construct more efficient index mechanism to complete fast retrieval becomes a significant problem.In recent years, with the advent of visual hash learning technologies, various methods including learning-based, multimodal-based and multitable-based have been proposed, which provides a new perspectives for hashing research. However, when dealing with massive datasets, most existing hashing methods still have many limits. Thus, this paper summarizes the previous classical work, and does further analysis about the advantages and disadvantages of the existing methods. The main work of this paper can be concluded as follows:First, we propose an adaptive multi-bit quantization approach for hashing. This method considers that most hashing methods usually adopt single-bit or double-bit quantization strategies, which destroy the data structure heavily and make related neighboring points close to the hashing threshold to completely different bits. Double-bit quantization neglects the fact that different projected dimensions have different amount of information, and yet adopts same number of bits for different projected dimensions. To cope with this problem, we propose an adaptive multi-bit quantization for hashing which utilizes data variances to measure the amount of information of different projected dimensions, i.e., the projected dimensions with larger variances should be allocated more bits, and the ones with smaller variances should be assigned single bit or zero bit. The learning process of hash codes is conducted with incomplete encoding and clustering strategy. The binary codes learned in this way can better preserve the neighbor structure and be quantified for fast, accuracy retrieval tasks.Second, we propose an anchor-based global similarity preserving hashing method. This method considers that most existing anchor-based hashing methods neglect the global semantic information of the data and only learn the binary codes in a local manner. Hence, we build a semantic label matrix and an anchor-based affinity graph, respectively. Then, we construct a local and global similarity preserving objective function. Our approach implements a more effective hash learning from two complementary, which describes the data structure more powerful than most existing methods.Experiments show that, compared with the classical hashing methods, the proposed approaches can achieve relatively high retrieval accuracy and scalability.