Discriminative affine local feature context based image classification

Most of context based methods focus on using context information at the visual word level without considering the relationship between local features. Besides, images are often captured with various poses, scale changes, illumination variation and camera parameters. This hinders the improvement of image classification performance. By combining contextual information of local features, this paper proposes a novel discriminative affine local feature context method for efficient image classification. We use the local feature at the position as well as other local features based on their distances and angels to this position. Affine transformations are done to the local feature context in order to get more robust and effective features. The discriminative affine-transformed local feature context is then chosen by minimizing the reconstruction error. Classification experiments demonstrate the effectiveness of the proposed method.