Abstract
Automatic understanding of visual information is one of the main requirements for a complete artificial intelligence system and an essential component of autonomous robots. State-of-the-art image recognition approaches are based on different local descriptors, each capturing some properties of the image such as intensity, color and texture. Each set of local descriptors is represented by a codebook and gives rise to a separate feature channel. For classification the feature channels are combined by using multiple kernel learning (MKL), early fusion or classifier level fusion approaches. Due to the importance of complementary information in fusion techniques, there is an increasing demand for diverse feature channels. The first part of the thesis focuses on the ways to encode information from images that is complementary to the state-of-the-art local features. To address this issue we present a novel image representation which can encode the structure of an object and propose three descriptors based on this representation. In the state-of-the-art recognition system the kernels are often computed independently of each other and thus may be highly informative yet redundant. Proper selection and fusion of the kernels is, therefore, crucial to maximize the performance and to address the efficiency issues in visual recognition applications. We address this issue in second part of the thesis where, we propose novel techniques to fuse feature channels for object and pattern recognition. We present an extensive evaluation of the fusion methods on four object recognition datasets and achieve state-of-the-art results on all of them. We also present results on four bioinformatics datasets to demonstrate that the proposed fusion methods work for a variety of pattern recognition problems, provided that we have multiple feature channels.