Bag-of-visual-words Model for Image Classification Based on Spatial Semantic Distribution

Abstract

To satisfy the requirement of image classification in the application of image retrieval, a novel method of image representation based on bag-of-visual-words model is proposed in the paper to describe the spatial semantic distribution of associated features. Firstly, the extracted SIFT features are mapped into visual words including certain semantic information. According to spatial pyramid hierarchy, the specific region is divided with local features, and the spatial distribution of associated features is analyzed from different aspects and in various regions. In this way, the semantic phrases are established with local features. Next, the spatial semantic lexicon is constructed with sparse encoding of spatial semantic phrases, and the images are described with the form of sparse statistical histogram vectors. Finally, the vectors of images are classified with the classifier embedded with the improved bag-of-visual-words model. The experimental results show that the accuracy of image classification is significantly enhanced which is benefited from the Bag-of-visual-words model with spatial semantic distribution.