الفهرس | Only 14 pages are availabe for public view |
Abstract In computer vision, the term shape refers to a variety objects such as images, videos and three-dimensional models that are collected from laser scans. Shape deformations refer to all kind of methods that aim to alter the object’s shape or form based on specific applications. The contentaware shape deformation is the process that depends on the local content/information of shapes in order to apply the required change. Segmentation is one of the most common problems that is considered an important pre-process step for a lot of computer vision applications. It aims at partitioning the object into multiple meaningful regions/segments in order to separately deal with each region or cluster. Concerning the image segmentation process, it depends mainly on the local content of image features such as color, boundary, texture, edges or any combination of attributes to locate objects in the image. In other words, each of the pixels in an image region is assigned the same label and is similar with respect to some characteristic or computed property. Graph-based segmentation methods are very promising and more appropriate to find exact solutions to solve the image segmentation problem via optimization methods. One of the most powerful graph based image segmentation methods is the GrabCut technique that depends on a probabilistic model in order to segment color images iteratively. One of the main contributions of this thesis is to apply modifications and extensions to the current semi-automatic binary-label GrabCut technique in order to solve existing problems and improve segmentation accuracy of natural images. All the proposed image segmentation techniques are evaluated for relevant accuracy criteria, and comparative studies are constructed with relevant techniques. II The semi-automatic GrabCut capabilities are extended to segmentation of human faces from images of full humans. The main contribution is the introduction of a new prior face location model to the GrabCut energy minimization function in addition to the existing color model. The location model considers the distance distribution of the pixels from the silhouette boundary of a fitted head, of a 3D morphable model, to the image. The proposed technique succeeds in eliminating the camouflage problem associated with the original GrabCut technique. In addition, it improves the segmentation accuracy with an error rate of 0.19% in comparison to the rate of 0.29% of the original GrabCut. Extension to the semi-automatic GrabCut aims at replacing the manual initial user intervention with the segmentation process with a fully automatic scheme that can segment images directly without any user guidance. Unsupervised image clustering techniques are considered an ideal solution for the automation process. The Orchard and Bouman (O&B), Self-Organizing Feature Map (SOFM) and Fuzzy C-Means (FCM) clustering techniques are selected for the development of the proposed clustering-based automatic GrabCut. A hybrid technique that combines advantages of using O&B-based and SOFM-based succeeds in producing the best segmentation accuracy with an error rate of 3.91% and overlap score rate of 90.16% in comparison to 5.49% and 87.71% rates achieved by original GrabCut. The RGB color space produced the best segmentation results with error rates of 4.25% and 5.4% using the O&Bbased and SOFM-based techniques respectively. The YUV color space produced the best results using the FCM-based technique with an error rate of 6.2%. Another contribution is the extension of the automatic binary-label GrabCut into a multi-label segmentation technique that can segment a III given image to its natural objects. The novel contribution provides the optimal solution via multiple piecewise iterative behavior instead of solving the NP-hard multi-labeling problem. On a dataset of 203 images, the O&B-based multi-label GrabCut achieves a ratio of more than 90% of the images with acceptable accuracy. Meanwhile, the SOFM-based multilabel GrabCut achieves a ratio of more than 94% of the images with acceptable consistency with human segmentations. Concerning segmentation of 3D meshes, it consists of subdividing a polygonal surface into patches of uniform properties either from a geometrical point of view or from a perceptual / semantic point of view. The other contribution of this thesis is to extend the use of the unsupervised clustering techniques to the problem of 3D mesh segmentation. The K-means and the FCM clustering techniques are selected for the development of the proposed clustering-based 3D mesh segmentation techniques. Based on empirical results on a dataset of 3D mesh models, the FCM-based mesh segmentation technique outperforms the K-means-based one in terms of accuracy and consistency with human segmentations. |