New Method Of Shape Analysis And Its Applications

Shape analysis is an important topic in computer vision and has won wide applications such as object recognition, content based image retrieval, OCR, medical diagnosis and so on. In this paper, several novel shape analysis methods are proposed and are applied in recognition of remotely-sensed image and plant-shape retrieval. The main work of this paper are as follows: (1) A hybrid genetic algorithm combined with split and merge technique is proposed to solve two types of polygonal approximation problems. (2) An invariant shape representation, interior angle chain, is proposed. (3) A novel Fourier descriptors derived from multi-level chord length function is proposed for shape retrieval.Polygonal approximation is an important shape representation method. However, acquiring a polygonal approximation of a boundary contour is a difficult task. There are two polygonal approximation problems which have attracted many researcher's attention. One is that given a digital curve, approximate it by a polygon with a given number of sides so that the approximation error is minimized. Another one is that given a digital curve, approximate it by a polygon with the minimum number of sides so that the approximation error does not exceed a given tolerance error. Traditional methods such as split and merge technique for polygonal approximation problem has a fast computational speed, but the quality of solution depend heavily on the starting point or the given initial solution. Although many global search technique based method such as genetic algorithm, ant colony optimization algorithm and so on can improve the quality of solution in some extent, the computational speel is very slower and the quality of solution is still not ideal and they can solve only one type of polygonal approximation problem. In this paper, a hybrid genetic algorithm is proposed for two types of polygonal approximation problems. Consider the difficulty of disposing the infeasible solution and the poor local search ability, we proposed a chromosome repairing scheme combined with traditional split and merge technique to dispose the infeasible solution. The advantage of this scheme is that an infeasible solution can not only be repaired but can also be pushed into a local optimal location in solution space. Experimental results and comparisons with other relative work show the superiority of our method. The proposed method has also been successfully applied to the polygonal approximation of contour of the lake in map.The second work is that an invariant shape representation, Interior angle chain (IAC), is proposed. The main idea is that an object's contour is firstly approximated by a equilateral polygon, the interior angle chain of the equilateral polygon is then calculated todescribe the shape.The similarity between two shapes can be measured by their IACs. The main contributions of our work are as follows. (1) An invariant shape representation is proposed and its invariance has been proved in theory. It is noted that the invariance doesn't require additional normalizing against the shape descriptor. (2) IAC reduces the number of dimension of the shape descriptor. The experimental results show its superior performance. IAC has also been applied to lake recognition in SAR images and all shown good results.The third work is that a novel Fourier descriptors derived from multi-level chord length function is proposed. Fourier descriptors is an important shape representation method and has won wide applications. Its main advantages are as follows: (1) it can remove the noise of the shape representation. (2) it is an impact descriptor. (3) it is very easy to be normalized to achieve invariance to scaling, translation, rotation. However, Fourier descriptors is affected by the contour function which is derived from. While the contour function exists the disadvantage of not considering all the shape information including the global shape information and the detail information for characterizing the shape. Considering the above problem, we proposed a novel contour function, multi-level contour function, which is obtain by equal-arc-length partitioning the contour. The contour function can not only characterize the global information but also consider the detail information. Its computation is also very simple. We derive the fourier descriptors from the multi-level contour function for achieving invariance to translation, scale, rotation and starting point. We use a large benchmark containing 1400 shapes to test the performance of MCLFD. The experimental shows that the average precision of MCLFD achieves 10 percentage higher that the other Fourier descriptors. We also applied it to the plant shape retrieval and the all results show that MCLFD is a promising shape descriptor.