In this paper, we address some new approaches for feature extraction, feature matching and feature grouping. These subjects are common and important in computer vision according to the theory frame-work of Marr. We first address a direct approach for designing differential operator, which is the most common method for feature extraction. Traditionally, to design a N-th order derivative filter (NODF), an Analytic Smooth Function Derivable up to N-th order (ASFDN) must first be found, whose N-th order derivative is the NODF. Our new approach allows us to design a NODF without using the ASFDN. This is important because we can find a number of NODF satisfied certain desired optimization criteria but their corresponding ASFDN may not exist. We first propose the sufficient and necessary conditions for a filter to be NODF, then the systematic design methods are addressed. New derivative filters are presented and the experiments shows that the new filter works better than the traditional filters. In the following chapter, we address the problem of feature points correspondence, which is a key step is stereo vision. We propose a new approach, the Reactive Tabu Search (RTS) approach, to this problem. Tabu Search is a mctaheuristic search technique that guides a local heuristic search procedure to explore the solution space beyond local optimality. Using RTS to minimizing a proposed cost function, we match the feature points, discard the outliers and recover the epipolar geometry in one step. Experiments on real images show that this approach is effective and fast. Chapt 4 deals with the feature grouping. The feature detector typically produces points or short linear disjointed edge segments, which by themselves are generally of little use unless they are grouped to form a higher level representation. We deal with two typical problems: geometric primitive extraction and linear feature extraction. Geometric primitive extraction is essentially an optimization problem. We propose Tabu Search technique to this problem. Our experiments show this approach works very well in extracting the line, circle and ellipse from the images. The second part of Chapter 4 deals with the linear feature extraction: road extraction from the satellite images. Humans can easily extract roads from images since humans can find the global saliency of roads. How to represent global saliency has been a challenging problem. Here, we propose an energy functional to represent this saliency. Based on the proposed energy functional, we proposed a new and efficient approach for road tracking and extraction. The experiments show that our proposed approach is efficient and without heuristic search.
修改评论