Research On Theory Of Structure Representation Of Object Contour Extraction

As computer technique and multimedia literature develop fast, being a kind ofinformation carrier, the images become more and more important and popular since theway it represents the world is easy, reliable, compact and direct. Though the tasks ofimage segmentation, object detection, boundary extraction(location or representation)have being researched for about three decades, they are still confronted with manyissues for they are essentially"ill posed", specially according to various situations andrequirements of every origins. Within these issues, particularly among the literature ofactive contour, the key problems of local minima, initialization sensitivity and e?ciencyin e?ectively object detection are still challenging, and awaiting for new strategies tosolve them.This dissertation tends to handle the problems and shortcomings of conventionalalgorithms, and to propose new approaches. It is noted that the task of boundary ex-traction is to obtain an appropriate way of representing the boundary either by curvesor discrete set of boundary points. For this, firstly, boundary extraction is achievedutilizing the powerful classification and approximation ability of neural networks, in-cluding radial basis function (RBF) network and SOM-like network ; secondly, byintegrating a variation of particle swarm optimization (PSO) algorithm into activecontour modeling, a kind of e?cient and intelligent active contour to locate the trueboundaries of the objects is obtained; thirdly, a new family of circular arc spline isproposed followed by establishment of a new energy function. By optimization thisfunction, the boundaries are interpolated and represented by series of consecutive arcs.In the end, a new vector field with swirling components is constructed, and then theproblem of boundary location is switched successfully into the problem of limit cyclelocation.Concretely, the contributions of the dissertation are described as follows:1. By establishing a compact RBF network, the boundary points are classified fromthe image domain. Centers, widths, and weights are three mainly considered fac-tors in constructing a RBF network. We propose using the triangular inequalitiesto lower computational cost of clustering the training samples; and propose amaximum degree spanning tree (MDST) to refine the coarse clustering (centers); then propose learning weights and refined widths using an anisotropic gradientdescent method. In the end, we can construct a compact RBF network with thethree obtained factors. Then we apply the network to object boundary detectionand gain good classification results in locating inhomogeneous boundaries.2. We propose detecting some feature points from the desired boundaries by a seriesof active circles with adaptive centers and radius, then clustering the featurepoints to obtain the corresponding RBF network. Using the inputs of the pixelintensities and gradient magnitudes, boundary detection is achieved in the casesof multiple boundaries, complex background, and noise and blurred boundaries.3. A coarse-to-fine and heuristic method of boundary location by a"SOM-like"net-work is proposed. Inspired from the traditional SOM network and universal grav-itation principle, this dissertation gives a coarse-to-fine boundary location algo-rithm. For each neurons, a union action is defined as the evolving direction andthe evolving-rate are adaptively changed by the referred gradients and positiongradients. With the changing union action and evolving-rate, the neurons evolveto the desired boundaries with appropriate manners. By generation updates ofthe feature points and the neurons, and multi-round evolution, the neurons willlocate the boundaries in a coarse-to-fine way. The algorithm performs well withinitialization robustness in complex long concavities, inhomogeneous and weakboundary location.4. Addressing to active contour modeling (ACM), i.e., SOM-ACM, this dissertationproposes a VBCPSO-ACM algorithm by treating the control points on the activecontour as the particles of an improved particle swarm optimization algorithm.In this algorithm, the particles's velocities are restricted in the correspondingdefined vector bundles, and depend on the best positions of other particles as well.By particle addition and deletion on the active contour, the active contour canapproach the desired boundaries in any precision. The algorithm can avoid self-intersection during contour evolving and also extract inhomogeneous boundaries.5. A novel method to reconstruct object boundaries from limited number of sparseand unorganized feature points with geodesic circular arc is proposed in this dis-sertation. A general form for a family of parametric circular arc spline is firstlyderived and followed by a novel method of arranging these feature points byminimizing an energy function depending on the circular arc spline configuration defined on the feature points from the desired boundaries. Considering the energyfunction is usually non-convex and non-di?erentiable, an improved scheme of par-ticle swarm optimizer is given to find its minimum. With this improved scheme,each pair of neighboring feature points along the boundaries arranged and thecorresponding directional chord tangent angles are computed simultaneously tofinish interpolation. It is shown experimentally and comparatively that the pro-posed method can perform e?ectively to restrict leakage on weak boundaries andpremature convergence on long concave boundaries. Besides, it has good noiserobustness and can as well extract multiple and open boundaries.6. Considering the"equilibrium issues"of external force field methods, in which thecapture range of the active contour is limited and the active contour is sensitiveto initialization. The task of object segmentation is novelly translated into limitcycle location in a defined vector field in the view of dynamical systems. An in-terpolated swirling and attracting ?ow (ISAF) vector field is firstly generated forthe observed image. Then, the states on the limit cycles of the ISAF are locatedby the convergence of Newton-Raphson sequences on the given Poincar′e sections.Consequently, with the computed periods of limit cycles, the objects'boundariesare represented by integral equations with the corresponding converged states.Experiments and comparisons with some traditional external force field methodsare done to exhibit the superiorities of the proposed method in the cases of com-plex concave boundary segmentation, multiple object segmentation, initialization?exibility. What is more, it is more computationally e?cient than traditionalACMs by solving the problem in some lower-dimensional subspace without usinglevel set methods.In conclusion, this dissertation proposes a series of approaches to addressing thekey issues in object boundary extraction, and establish the methods of structure rep-resentation for object boundary extraction models to make boundary extraction bedone e?ectively.