Experimental study on segmentation methods in road recognition

摘要

The digital image processing is extremely important for numerous areas as a significant one is Earth observation. The identification of the land cover using satellite images is very important for most of the economics spheres. The image segmentation is known as a basic option for the process of classification. It works as an improving element for the performance and for the accuracy. The main role behind the image processing is to provide the recognition of the shapes and objects in an image. In this process a segment has a significant role. A segment is actually a homogenous part of any image. The survey of image processing applications shows examining, refining and combining of already outlined segments are featured. The delineation of the segments is momentous, too, when it comes to the quality of the results. The main goal behind this article is to make a comparison between effectiveness of several graph-based image segmentation algorithms in segmenting of the roads. These are the best merge algorithm of Beaulieu, Goldberg and Tilton, the tree merge segmentation of Felzenszwalb, the minimum mean cut segmentation of Wang and Siskind and the normalized cut algorithm of Shi and Malik. The represented methods in this article are used in segmentation of orthophoto image of an urbanized zone including roads. The image is determined as a matrix of pixels, while they are also vectors of the intensity numbers, which are usually registered by the remote sensing sensors. The results from the experiments are shown, discussed and lead to following conclusions. The tree-merge segmentation by Felzenszwalb and Huttenlocher is not suitable in the road segmentation. Thematic preciseness of the normalized cut segmentation by Shi and Malik is not shown the needed accuracy in this experiment. The best merge method by Tilton shows the most satisfying indexes.