Automatic image annotation method based on cross-media sparse theme code

Abstract

The invention relates to an automatic image annotation method based on a cross-media sparse theme code. The method comprises the steps that firstly an image word is generated by an image in a multi-media document, a word bag model is used for representing an annotation word in the multi-media document as a vector, and the processed multi-media document is obtained; according to the processed multi-media document and a probability theme model, an image cross-media sparse theme code model is obtained; by using a maximum posterior probability estimation method, the image word and the annotation word in the multi-media document and a joint distribution formula of a relation code variate of the image word and the annotation word are obtained, and an image word code, a multi-media document codeand a relation code in the joint distribution formula are modeled by adopting Laplace transcendence and super-Gaussian; a coordinate descent method is used for optimally figuring out a cross-media sparse theme code model, and then the cosine similarity degree between an image code and an annotation word code is calculated and subjected to image annotation. According to the automatic image annotation method based on cross-media sparse theme coding, the complexity of annotation time and space is lowered, the accuracy rate of image annotation is guaranteed, and meanwhile the efficiency is guaranteed.

Description

technical field [0001] The invention belongs to the field of computer application, image processing and data mining, and in particular relates to an automatic image labeling method based on cross-media sparse topic coding. Background technique [0002] With the development of the Internet and the popularization of digital devices, the amount of image resource data has grown exponentially. How to effectively retrieve and manage image resources is an important topic in the field of computer vision. Traditional content-based image retrieval uses low-level image features, which cannot establish a good relationship with high-level semantics, that is, there is a semantic gap, which makes retrieval quality difficult to meet the requirements. The automatic annotation of images is to establish the mapping or association relationship between the image visual feature space and the high-level semantic space through known images, that is, to project the two heterogeneous media data of a...

AI Technical Summary

Problems solved by technology

Although the probabilistic topic model achieves superior performance for image annotation, there are two limitations: the probabilistic topic model cannot effectively control the sparsity of the latent representation; This method is not suitable for scenarios where a word is associated with multiple images

Images