Image Segmentation Techniques

Segmentation is one of the most difficult and important steps in digital image processing. Segmentation accuracy determines the eventual success or failure of computerized analysis procedures. For this reason, image segmentation has been widely investigated for more than 40 years, and hundreds of algorithms have been presented in the literature. Although those algorithms are to some extent successful, image segmentation is still far from been solved. There is no single method which can be considered good for all images, nor are all methods equally good for a particular type of image. Moreover, due to the lack of systematic theory, there is no idea to guide us how to choose appropriate algorithms for different images. Even a universally accepted technique for evaluation of segmentation results is not existed. As a result, it is very difficult for us to find suitable algorithms when facing a practical image segmentation problem. The algorithms, though delicately designed, can only be successfully applied to a part of images. Therefore, several different segmentation approaches are investigated in this paper, and machine-learning techniques are also used to meet the challenge of automated image segmentation.According to their different characteristics, images can be roughly separated into two classes: textures and non-textures. Non-texture images usually have simple content, and can be properly segmented by gray level threshold. In traditional thresholding algorithms, it is determined only by its own gray level that a pixel should be classified into background or target. Therefore, when applying a thresholding algorithm to an image, we generally assume that the background and target in that image occupy different gray level zone. However, that assumption is not hold well in most cases because of the impact of noise and unbalanced illumination. In this paper, the strategy of combining the information of gray level and spatial position is adopted to achieve non-texture image segmentation, and two novel approaches are presented. The first approach is a generalization of Otsu algorithm. Based on the gray level and average gray level scatter-plot, an approximate threshold is estimated by maximizing the variance between background and target. Then the image is roughly segmented into background, target and transition by using that threshold. Finally, the transitional pixels are further classified by employing region growing technique. The second approach aims to segment Synthetic Aperture Radar (SAR) image. Different from the former one, cluster method is used to obtain coarse segmentation in fuzzy domain. Those two approaches achieve both robustness and less computation cost.As an important characteristic for the analysis of digital images, texture has always attracted attentions from researchers in image segmentation society. Consequently, quite a lot of methods have been developed to describe it. This paper mainly discussed the fractal based texture descriptors and their application in segmenting remote sensing images. Taking into account of the facts that most widely used fractal estimations produce less accurate resultsand traditional morphological method suffered from heavy time expense, we present a modified morphological estimation of fractal dimension, which can give more accurate result at an acceptable computation cost. Furthermore, we generalize above method to estimate multifractal model, and thus proposed a novel set of multifractal features: local morphological multifractal exponents. Using those features, texture segmentation algorithms are presented and tested on natural images.Image segmentation is easy for human beings, but quite difficult for even the most powerful computers. We attempt to exploit various machine-learning techniques to learn the heuristic knowledge from users' experiences, so that the image segmentation system can have some human ability in adaptively selecting optimal algorithm and corresponding parameters. Learning based image segmentation system can be classified i...