A soft relevance method for content-based scene categorization in the BoW framework

A new soft relevance technique for scene categorization is proposed in this paper. A popular approach for scene categorization is the Bag-of-Words (BoW) framework, where a histogram is calculated for each image as the image signature. However, in most of the existing BoW based image classification methods, all the image signatures are regarded equally, so the outlier images may be harmful to the classification performance. In view of this, this work tries to address the issue by estimating the soft relevance value of the image signatures using image signature space modeling and then incorporate it in Fuzzy Support Vector Machine (FSVM). The effectiveness of the proposed method is validated on NTU Scene-25 dataset, and it is shown to outperform some state-of-the-art methods in BoW framework for scene categorization.