Research On Hash Ranking For Large-Scale Image Retrieval

Since the 21 st century,with the growing popularity of a variety of digital camera equipment and image processing tools,massive image data has entered into people's lives.In reality,the demand for image retrieval and image recognition makes retrieval of image based on image content in a large image library very important.In order to improve the efficiency of image retrieval,the researchers map the image as a binary hash code by the hash method,and measure the similarity degree of the image according to the Hamming distance between the corresponding hash codes.This method has the advantages of compressed storage and fast query speed,but there are still shortcomings.The quantization in the hashing process usually degenerates its discriminative power when using Hamming distance ranking.The number of images that share the same Hamming distance with the query image is large,which will result in the inability to rank these images effectively and reduce the performance of the search.Therefore,it is important to perform a more fine-grained ranking of image.Hash bit weighting has proven to be a promising solution for fine-grained ranking.In this paper,we use local neighbor relation matrix between data points and anchor points to effectively approximate the original similarity between data points based on the previous study of the anchor graph.This method can overcome the shortcomings of the standard Euclidean measure which can not strictly capture the global similarity between data points,and can obtain the similarity between samples in any data set.The similarity between the query point and the database sample is constructed by the anchor graph.On this basis,we use the discriminative power of each hash function and their complement for approximate nearest neighbor search to learn a set of query adaptive bit wise weights.And then we propose a new hash ranking method based on the weighted Hamming distance.Assigning different weights to the hash bits to distinguish the returned results with the same Hamming distance,resulting in a finer and more accurate ranking order.This method is a general weighting method for all kinds of hashing algorithms without any strict requirement for data distribution.Finally,according to the proposed method,two data sets,MNIST and NUS-WIDE which are recognized by image retrieval areas,are tested.According to the performance evaluation index,the experimental results show that the proposed method is more general and effective than the existing methods.