Study Of Image Segmentation Algorithm Based On Fuzzy Connectedness

Image segmentation is the key to the image processing, the results of that determines thefinal quality of the image analysis and pattern recognition. In recent years, many academichave proposed various methods to achieve it. All this methods that can be divided into basedon region or border. Fuzzy connectedness (FC) algorithm is a method based on region. First,by calculating the similarity between the pixels to build the network topology; then, findingoptimal path between seed points and others points through the network; finally, computed theFC of all points in the image.After summarize some of the related problems about the FC algorithm, this paper madesome improvements:Firstly, for the problem that objects extraction. Artificial select the threshold increases thenumber of interactions, and cannot accurately determine the optimal threshold. Relative fuzzyconnectedness (RFC) algorithm needs to count twice FC, doubling of computing time. Wehave presents combined the Otsu algorithm with FC to automatic threshold selection. Theexperiments show that this method is not only faster than the RFC and more accuratesegmentation results when the background is complex.Secondly, inoder to find the optimal path, a method that based on Dijkstra’s greedy algori-thm to improve the speed is present. Each time to find the global optimal solution, through aunified root fuzzy connectedness way to remove the update of the original algorithm, eachpixel is calculated once a vague connection to complete the degree. Meanwhile, each time‘siteration may be implemented to connect the plurality of pixels fuzzy calculation, reducing thenumber of iterations. By real medical image tests show that compared with other improvedalgorithm, which split significantly faster, and does not affect its segmentation accuracy.Thirdly, in order to make the segmentation results more accurate and more stable, and hasnoise immunity, introduce scale to express the local structure properties. But there is somedefect, such as the calculation of the scale is time-consuming, sensitive to noise and threshold.For these problems, a new method for calculating the scale is present. First to detect the edgesof the image use of Canny operator, and then design an adaptive template at the current point,looking for the edge points from the template, the sacel is the distance from the current pointto the nearest edge point in the template. By the chest and simulating brain images segmentati-on proved that the new scale is faster than the previously, and almost no effect the results,moreover the new scale is better than the original in the strong noise image.Finally, tensor scale operator can display the directionality and anisotropy of localstructures features that scale operator’s drawback. However, the tensor scale operator iscomputationally intensive and there are some error rates. This article will introduce anon-linear structure tensor to express local properties, the experiments show that under thesame conditions, the FC algorithm based on structure tensor is speed up much times than theFC based on tensor scale, and the accuracy of segmentation is hardly affected.