| Summary: | 碩士 === 國立中興大學 === 資訊管理學系所 === 99 === Image segmentation is an important technique in image content understanding and pattern recognition. Nowadays, many of image segmentation techniques have been proposed, but none can segment all kinds of images well enough. Traditional image segmentation methods usually cause the results to be to be over-segmentation. For example, if the sky shows strong continuous variations of color, it is easily segmented into many regions. The purpose of this thesis aims to solve this problem.
Two region merging methods are proposed to improve the over-segmentation problem produced by traditional image segmentation algorithms. By applying region merging, the regions of over-segmentation are reduced gradually to export final segmentation. In the first method, over-segmentation structure is built as region adjacency graph. A special vertex selection algorithm is employed to choose the vertex with the smallest region size, thus there may exist two or more vertices that satisfy the condition. These vertices and their neighbors are being examined to determine their similarity. The vertices are merged at the end of each iteration if they are found similar. The verification method is to calculate inter-class difference and extra-class difference between two vertices in L*, a*, b* channels. If the inter-class difference of two vertices is smaller than the extra-class difference in all three channels, two regions should be merged. There may occurs regions merged at each iteration. If merging criterion is reached, the algorithm would find the iteration number which gives the best segmentation, and then reruns and stops at that iteration. Finally, the final segmentation is exported. In the second method, the algorithm sifts some smaller edges by using minimum spanning tree. Histograms of three color channels are divided into different parts, respectively; each part is labeled as a number, and each vertex has a set of color code formed by these label numbers. If the color codes of two regions satisfy certain conditions, two regions would merge at the end of iteration. The algorithm stops when the number of regions in the image does not reduce. Experimental results showed that the proposed methods can improve the problem produced by traditional algorithms.
|
|---|
