Please use this identifier to cite or link to this item:
|Title:||Shape analysis using multiresolution gradient vector field||Authors:||Goh, Wooi Boon||Keywords:||DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision||Issue Date:||2006||Source:||Goh, W. B. (2006). Shape analysis using multiresolution gradient vector field. Doctoral thesis, Nanyang Technological University, Singapore.||Abstract:||It is widely held that our perception of visual form is due to the excitation of cells sensitive to orientation and edge information at the boundary of the object. It is hypothesised in this work that the multiscale spatial integration along the hierarchical visual pathway do in some way migrate boundary-based orientation into the interior region of the shape and this dense multiresolution gradient vector field (MGVF) can be exploited for robust representation of shapes. Based on this premise, the thesis proposes a novel computationally efficient multiresolution framework for generating a dense gradient vector field for the interior region of a silhouette shape. Within this framework, the smoothness of the MGVF can be easily adjusted via a single smoothing parameter. The thesis subsequently explores the robustness and usefulness of the MGVF in two-dimensional shape analysis. Firstly, a multiresolution algorithm based on the detection of local directional disparity maximas in the gradient vector field is proposed for extracting the medial representation of a shape. It is shown that the MGVF skeleton can be extracted in a noise-robust manner and is invariant to both scaling and arbitrary rotation. Secondly, a global shape descriptor in the form of an orientation histogram is proposed to describe the gradient vector field characteristics within the shape. Comparative evaluation with other popular global descriptors suggests that this descriptor strikes a good compromise between boundary noise robustness, computational efficiency and the ability to classify shapes. Though compact and computationally efficient, the representation of an entire shape with a single global statistical feature made the descriptor sensitive to commonly occurring shape distortions such as occlusion, part articulation, deletion and addition. This finally led to the proposal of a part-based descriptor that incorporates both the MGVF skeleton and the orientation histogram. The decomposition of a shape into parts is based on the segmentation and merging of the MGVF skeletal segments. However, final decomposition is deferred till the shape matching stage as this thesis argues that shape decomposition is a dynamic process, which can only be properly carried out when two shapes are being matched. Comparative evaluations of the proposed MGVF part-based descriptor with those in recent literature show that its shape retrieval performance is more accurate.||URI:||https://hdl.handle.net/10356/2637||DOI:||10.32657/10356/2637||Rights:||Nanyang Technological University||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Theses|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.