A Hypergraph-Based Image Retrieval and Annotation Method

Abstract

The invention discloses a hypergraph-based image retrieval and tagging method. By utilizing a hypergraph theory, a comment relationship, a tagging relationship and a visual similarity relationship of social contact images are effectively organized; by utilizing batch, parallel and buffer technologies, the calculation efficiency of a transfer probability matrix of a hypergraph is improved; by utilizing upper and lower bound estimation and approximation methods for ranking scores of nodes, the query efficiency is improved; and by utilizing user feedback, the query quality is improved. An image t-NN graph is generated according to visual features of the images, a hypergraph model is built in combination with social correlation information of the images, and the transfer probability matrix is calculated and stored in a B+ tree; a parallel personalized PageRank query is performed, and candidate point sets are screened according to upper and lower bound estimation and are sorted; and finally according to the user feedback, the query is performed again, and a query result is optimized. Various multidimensional relationships can be effectively organized; the calculation efficiency and the query efficiency of the transfer probability matrix are greatly improved; the query quality is higher than that of an existing method; and the best performance is provided.

Description

technical field [0001] The invention relates to a Top-k query technology on a hypergraph, in particular to an image retrieval and labeling method based on a hypergraph. Background technique [0002] With the development of social media and mobile Internet, social image sites provide a large number of images marked with text by different users. Social images are often accompanied by a variety of information, such as visual features, tags, and users, as well as a variety of behavioral relationships, such as annotations and comments. Retrieval and labeling of massive social images is widely used and has become a research hotspot in the fields of database, data mining and machine learning. [0003] Image retrieval is to find the closest image object according to the given information. According to the different types of given information, there are many retrieval types such as similar image retrieval and keyword image retrieval. Image annotation is to add semantic text informa...

AI Technical Summary

Problems solved by technology

However, a single feature can only represent certain aspects of relevance, and cannot be used to represent real semantic associations. Ordinary graphs cannot represent high-dimensional relations (such as annotation and comment relations), resulting in information loss.

Some methods in the field of machine learning use hypergraph models, but use complex matrix operations to solve them, which have huge time and storage overhead and are not scalable

Images