A Multi-feature Fusion Method for Automatic Multi-label Image Annotation with Weighted Histogram Integral and Closure Regions Counting

Abstract

In order to reduce the semantics gap and improve the consistency between structural difference among images and similarity of semantics from the corner of cognitive, many approaches on automatic image annotation have been developed vigorously. In addition to making use of both the features of n-order color moment and texture information, a multi-feature fusion method for automatic image annotation was proposed by using weighted histogram integral and closure regions counting in this paper. Based on Corel image data set, it showed in our experiments that the proposed approach can achieve better performance than that of traditional one using multi-label learning k-nearest neighbor algorithm, i.e., it can improve average precision measure index from 0.222 to 0.352 in automatic image annotation.