Computational models for image contour grouping

Abstract

Contours are one dimensional curves which may correspond to meaningful entities such as object boundaries. Accurate contour detection will simplify many vision tasks such as object detection and image recognition. Due to the large variety of image content and contour topology, contours are often detected as edge fragments at first, followed by a second step known as {u0300}{u0300}contour grouping'' to connect them. Due to ambiguities in local image patches, contour grouping is essential for constructing globally coherent contour representation. This thesis aims to group contours so that they are consistent with human perception. We draw inspirations from Gestalt principles, which describe perceptual grouping ability of human vision system. In particular, our work is most relevant to the principles of closure, similarity, and past experiences. The first part of our contribution is a new computational model for contour closure. Most of existing contour grouping methods have focused on pixel-wise detection accuracy and ignored the psychological evidences for topological correctness. This chapter proposes a higher-order CRF model to achieve contour closure in the contour domain. We also propose an efficient inference method which is guaranteed to find integer solutions. Tested on the BSDS benchmark, our method achieves a superior contour grouping performance, comparable precision-recall curves, and more visually pleasant results. Our work makes progresses towards a better computational model of human perceptual grouping. The second part is an energy minimization framework for salient contour detection problem. Region cues such as color/texture homogeneity, and contour cues such as local contrast, are both useful for this task. In order to capture both kinds of cues in a joint energy function, topological consistency between both region and contour labels must be satisfied. Our technique makes use of the topological concept of winding numbers. By using a fast method for winding number computation, we find that a small number of linear constraints are sufficient for label consistency. Our method is instantiated by ratio-based energy functions. Due to cue integration, our method obtains improved results. User interaction can also be incorporated to further improve the results. The third part of our contribution is an efficient category-level image contour detector. The objective is to detect contours which most likely belong to a prescribed category. Our method, which is based on three levels of shape representation and non-parametric Bayesian learning, shows flexibility in learning from either human labeled edge images or unlabelled raw images. In both cases, our experiments obtain better contour detection results than competing methods. In addition, our training process is robust even with a considerable size of training samples. In contrast, state-of-the-art methods require more training samples, and often human interventions are required for new category training. Last but not least, in Chapter 7 we also show how to leverage contour information for symmetry detection. Our method is simple yet effective for detecting the symmetric axes of bilaterally symmetric objects in unsegmented natural scene images. Compared with methods based on feature points, our model can often produce better results for the images containing limited texture.