Scene classification based on category-specific representations created through prototype feature selection

• Authors: Shuang Bai;Tetsuya Matsumoto;Hiroaki Kudo;Noboru Ohnishi;Yoshinori Takeuchi
• Venue: Proceedings of the 27th Conference on Image and Vision Computing New Zealand
• Year: 2012

In this paper, we propose a novel approach to coping with scene classification. In natural scenes, images from different categories often share similar components. As a result, it is difficult to distinguish them directly. In order to overcome this problem, for an input image, we propose to create a set of category-specific representations and use them to model the image as a probability distribution over the categories. Specifically, we first create a prototype for each scene category by pooling local features of its sample images together. Then, based on the category prototypes, given an input image, its salient local features corresponding to each category are extracted and used for creating a category-specific representation, respectively. Here, salient features for a scene category are defined as features that appear frequently in its instance images. In the classifier training stage, corresponding to one category prototype, the category-specific representations of a set of training images are used to train a multi-class SVM classifier. Thereafter, the obtained SVM classifiers are used to classify another set of training images in a probabilistic way, so that for each category-specific representation of the second set of training images, a probability vector is able to be obtained. Subsequently, all probability vectors of a training image are concatenated as its final representation, which are used to train another SVM classifier. For an unknown image, the same process is applied to it for classification. The proposed method is evaluated on datasets scene categories 8 and scene categories 15, experiment results demonstrated the effectiveness of the proposed method.