Font Size: a A A

Research On Active Contour Models For Image Segmentation

Posted on:2014-03-11Degree:DoctorType:Dissertation
Country:ChinaCandidate:L M QinFull Text:PDF
GTID:1268330401971013Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
Snakes, or active contours, have been widely used in computer vision and image processing applications. An external force for snakes called gradient vector flow (GVF) which largely addresses traditional snake problems of initialization sensitivity and poor convergence to concavities has gained tremendous popularity, while generalized GVF (GGVF) aims to improve GVF snake convergence to long and thin indentations (LTIs). However, the shortcomings of GVF and GGVF snakes including convergence to LTI, edge preserving, computational cost, as well as noise, parameter, and initialization ro-bustness restrict their further applications. To solve these problems, this dissertation pro-poses two novel external forces and a parameter estimation approach for external forces as follows:This dissertation proposes a novel external force termed as component-normalized GGVF (CN-GGVF) to improve active contour convergence to LTIs. We find and show that GGVF snakes only have limited capability to converge to LTIs. We investigate thor-oughly the characteristics of GGVF external forces within LTIs, and accordingly identify and verify two cruxes of GGVF snake convergence to LTIs, referred to as noise and oblit-eration problems. We present a method for suppressing the noise problem, and propose a novel external force termed CN-GGVF to eliminate the obliteration problem. CN-GGVF external forces are obtained by normalizing each component of initial GGVF vectors with respect to its own magnitude. Experiments demonstrate that compared with GGVF snakes, the proposed CN-GGVF snakes are able to capture LTIs regardless of odd or even widths with a remarkably faster convergence speed, and achieve better performance on the real photographic image testing, while preserving other desirable properties of GGVF snakes with lower computational complexity in vector normalization.This dissertation proposes a novel external force termed as adaptive anisotropic G-GVF (AAGGVF) to solve the problems of convergence to LTIs, noise robustness, and edge preserving for existing active contour models. Based on local image features, AAG-GVF adaptively adjusts weighting coefficients of the smoothing and data terms, and si-multaneously adaptively adjusts the weighting coefficients of the diffusions along the normal and tangential directions of the isophotes. Experimental results demonstrate that compared with existing state-of-the-art snakes, the proposed AAGGVF snakes have an extended capture range, a superior noise robustness, the highest diffusion efficiency, and are simultaneously able to preserve weak edges under noisy environments and capture noisy LTIs regardless of odd or even widths with a remarkably faster convergence speed.This dissertation investigates in depth the relationship between the capture range of the desired features (like edges) and the parameters of the GGVF field, and accord-ingly proposes a parameter estimation method for calculating the desired GGVF field. We look into the shape and size of the capture range of a single GGVF vector, and ex-plicitly show that they are specifically determined by the discrete approximation of the two-dimensional (2-D) Laplacian operator and the iteration number, respectively. With the specific relationship, we investigate in detail and verify the capture range of the de-sired features is uniquely determined by the iteration number of external force field and present a parameter estimation method to calculate the desired GGVF field for the correct convergence of the snake in noise-free environments. In noisy environments, we divide the relationship between the desired features and the noise into four categories according to the strength and number of their gradient vectors. For each case we study in depth the relationship between the capture range of the desired features and the parameters of external force fields. Finally, we propose a parameter estimation method for calculat-ing the desired GGVF field in two cases. A series of image segmentation experiments on both synthetic and real images as well as object tracking experiments demonstrate the effectiveness, efficiency, and universality of the proposed capture range analyses and parameter estimation methods.
Keywords/Search Tags:Active contour model, external force, gradient vector flow, anisotropicdiffusion, capture range, image segmentation, object tracking
PDF Full Text Request
Related items