Complex Filters and Higher-Order Spatial Information for Image Categorization

Abstract

This Thesis applies complex spatial filters to the front end filtering to a computer vision framework for object recognition and scene categorization. This involves careful filter design in the Fourier domain based on discrete frame properties. Biological plausibility of the suggested filtering is compared against a common model found in the computer vision literature. The designed complex filter bank is equipped with focus-of-attention operators. Specifically, two possible keypoint detection methodologies are examined and compared with state of the art keypoint detection methods. This includes an investigation of scale-estimation methods. In addition, three image patch descriptor arrangements are proposed to sample the complex filter responses, and an initial evaluation of categorization performance is undertaken. Next, the spatial pooling arrangement of the best performing descriptor is further optimised and the performance of different complex filter bandwidths is examined in class separation tasks. A further study is conducted on the effects of a Winner-Take-All (WTA) approach to modifying filter responses before pooling. A thorough evaluation of descriptor performance is undertaken to reveal any advantages or disadvantages from a variety of perspectives. Next, the clustering behaviour of descriptors of various types is inspected in the descriptor feature space. A reverse look-up of visual words attempts to relate clustering behaviour to descriptor performance. Typical grouping approaches, such as spatial pyramids, are then compared with a novel method for coupling visual words in which a linear kernel SVM learns class separability. A final evaluation on this stage is presented and discussed, leading to conclusive arguments about the importance of careful approaches to word-pairing for good-quality categorization.