Utilizing Context Information to Enhance Content-Based Image Classification

Abstract

Traditional image classification relies on text information such as tags, which requires a lot of human effort to annotate them. Therefore, recent work focuses more on training the classifiers directly on visual features extracted from image content. The performance of content-based classification is improving steadily, but it is still far below users’ expectation. Moreover, in a web environment, HTML surrounding texts associated with images naturally serve as context information and are complementary to content information. This paper proposes a novel two-stage image classification framework that aims to improve the performance of content-based image classification by utilizing context information of web-based images. A new TF*IDF weighting scheme is proposed to extract discriminant textual features from HTML surrounding texts. Both content-based and context-based classifiers are built by applying multiple correspondence analysis (MCA). Experiments on web-based images from Microsoft Research Asia (MSRA-MM) dataset show that the proposed framework achieves promising results.