Image Hash retrieval method based on KPCA multi-table index

Abstract

The invention provides an image Hash retrieval method based on the KPCA multi-table index. The method includes the steps of distinguishing feature selection, feature clustering, multi-table index construction and Hash code optimization. According to the method, before a Hash projection function is trained, firstly, features with the distinguishing capability are selected from image feature dimensions through principal component analysis based on a kernel function to serve as a training set; on this basis, clustering centers of different semantic samples are acquired through a feature clustering method, and multiple optimal neighboring clusters of each cluster are found; finally, hierarchical division is conducted on a clustering space to construct multiple index tables. In retrieval, multiple Hash index tables are searched to improve the retrieval performance. In the method, the high-dimensional image features are mapped into simple binary codes, and thus the data storage space is saved; the problem that when a single-table index structure is adopted, similar images have a large dispersion difference or similar features have a large attribute distribution interval, in other words, original similar features are mapped to different Hash codes is solved.

Description

technical field [0001] The invention belongs to the field of image retrieval, relates to a content-based image retrieval method, and is suitable for large-scale image retrieval and nearest neighbor search of high-dimensional data. Background technique [0002] The image database management system in the 1970s manually marked the semantic content of the image, and used traditional database technology or text information retrieval technology to store and index the semantic keywords of similar images. Its advantage is that it is based on mature database retrieval technology and text content indexing technology, and the retrieval speed is relatively ideal. However, web pages are text information associated with images rather than feature information related to image content, resulting in partial index results that do not meet user requirements, and with the advent of the era of big data, the scale of image data has grown exponentially. This subjective and inconsistent manual la...

AI Technical Summary

Problems solved by technology

However, there must be two important problems in this way: 1) image features are often a kind of high-dimensional data, high-dimensional data storage requirements are high, and the computational efficiency and class-to-class distinction are low; 2) It is difficult to meet the efficiency requirements for linear search on large-scale data

[0006] However, existing methods need to solve a problem. Image features are encoded into binary hash codes through hash mapping functions. These mapping functions not only need to be highly discriminative, but also be able to distinguish multiple types of features, otherwise it is difficult to ensure the accuracy of retrieval. sex

Images