Visual Learning by Feature Combination and Feature Construction

In developing a visual learning method, the selection of features highly affects the performance of the method. However, the optimal features generally depend on learning tasks. Therefore, it is necessary for the effective learning to find the optimal features according to the learning task. In this paper, we propose two types of new visual learning methods; the feature combination method and the feature construction method. The former is the simpler method that combines some low-level features into a higher-level feature represented by a linear combination of the low-level features. This method can find more discriminative boundary in the feature space. The latter is the more complex method that combines a variety of low-level features into a constructed feature represented by a tree structure. Although this method is rather complex than the feature combination method, much more discriminative boundary can be found because such a novel feature can form the more discriminative feature space. We compare the performance of our methods based on some experiments with some real-world images.