A Skeleton-Based Algorithm For Object Segmentation

Image segmentation is a basic and important technique in image processing. It is also the premise and key to object recognition and image analysis. Image segmentation, which plays an important role in our daily life, is widely used in the fields such as face recognition, machine vision, medical image and pedestrian detection et al. But it is still a difficult problem for automatic segmentation of touched objects via a computer. Due to the difference of object structure and quality of the image, there is no universal segmentation solution that can be applied to all the image segmentation problems. In this paper, a new skeleton-based algorithm is proposed to segment touched rigid objects.In order to verify the proposed algorithm, one object segmentation and counting system is designed to realize segmentation and quantitative statistics of the objects. A camera with fisheye lens is used in the system to acquire images, but the acquired images have severe radial distortion which seriously affects later process. Therefore, how to solve the problem of image distortion is the first difficulty of the system. In this paper, a new method is proposed to solve it without nonlinear correction of the distorted image. Skeleton, which integrates the geometrical and topological features of objects, is a very useful shape descriptor and it is widely used in the field of pattern recognition. But skeleton is very sensitive to noise and deformation of the edge of objects, so skeleton extraction is the second difficulty of this paper. The medial axis transform algorithm based on Euclidean distance is used to extract skeletons of objects in this paper. Euclidean distance is obtained by a fast Euclidean distance transformation algorithm in two scans. Generally, there are both isolated objects and touched objects in the image. In order to segment the touched objects more effectively, isolated objects should be firstly extracted and removed from the image. Besides, after some touched objects are segmented, they should be determined whether or not isolated objects. So the determination of isolated objects is the third problem in this paper. A skeleton-based criterion is proposed to determine the isolated objects. By comparing skeletons of objects with the standard skeleton, isolated objects can be extracted from the image. Segmentation of touched objects is the last key problem of this paper, different solutions are applied to segment different touched objects. For slightly touched objects, the skeleton points, whose gray values are smaller than a threshold, are firstly removed from the skeletons; then slightly touched objects can be segmented by reconstruction of them with pruned skeletons. However, for heavily touched objects, the skeleton points with abnormal gray values or generated by the touch of objects are removed from the skeletons; then the pruned skeletons are extended; finally, take each point of extended skeletons as center and search segmentation points in each touched place, heavily touched objects can be segmented by setting the pixels on the line between two matched segmentation points as background. The experimental results show that different touched objects can be effectively segmented by the proposed algorithm.