Image Segmentation Based On Graph Theory And Probability Statistics

This paper primarily studies the segmentation algorithm of natural images and pavement cracks images. Due to the existence of blurring problem between the boundaries of foreground and background or target areas in natural images, and how to obtain a well segmentation results to be difficulty lies. Besides, the uneven illumination, complex background and lots of noise in pavement cracks images are equally presented difficulties to segment. Since the complexity of the image, using a single image processing algorithm is difficult to obtain desired results, so requiring combination with other segmentation algorithms.The paper describes the basic principles of graph theory applied to image segmentation, using the principle of minimum spanning tree of graph theory for image segmentation. First, the image is mapped into a graph, then segmenting it based on the mature theory of graph theory and other algorithms. The main contributions and innovations are as follows:(1) Bayesian matting algorithm can separate foreground from background, and deal the problem of fuzzy edges well, so it will be as preprocessing algorithm of natural image. Experiments show that the results of bayesian matting are better than others. Segmenting images based on graph theory after matting can not only avoid misclassification of foreground and background, but also keep details. In the same time, processing the foreground and background use improved graph based method separately has an advantage of setting different segmentation scales according to their texture and color characteristics, which avoids over-merging and over-segmentation in a certain extent.(2) For the minimum spanning tree algorithm, large parameters lead to over-merging, and small parameters result in over-segmentation, it is difficult to get a certain segmentation scale. For some images, it is easily affected by noise and isolated points as the classical method only considers the maximum weight of intra-region while the weight function not take the distance of spatial location into consideration, so the graph-based algorithm is improved in three main aspects:the function of intra-regional and inter-regional differences; the function of edge weight; and re-merge mechanism after segmentation in graph mapping. The improved algorithm reduces the impact of noise and experiments show that it can increase segmentation accuracy by 6%-12% effectively.(3) As the images of pavement cracks have dimly lit, weak boundary and lots of noise, a guided filtering based enhancement algorithm is proposed. Compare to classical algorithms, it not only can improve contrast between cracks and the background, keep edge information, but also has high efficiency. According to the experiments, the guided filtering has good results in de-noising, detail smoothing and enhancement.(4) We use the minimum spanning tree based method to process the enhanced cracks images with K-means clustering. The algorithm is simple and efficient, both considers local features and global information of the image. Using skeleton extraction, burrs remove and broken connections of mathematical morphology to be post-processing algorithms. A number of experiments show that the algorithm can achieve better segmentation results for most of pavement cracks images.