A 3 Levels Image Segmentation with Integrated Objects and Local Information

This paper introduces a new integrated method for segmenting image which imitates the human observing mechanism. Human firstly pays attention to large objects, then some details, and lastly the concerned details. At the first stage of the algorithm discontinuous edges are detected in the original image, then cut equidistantly to some segmentations, and the end points of segmentations are gathered into a points set. At the second stage a constrained Delaunay triangulation is yielded from the points set and then a graph is attained too. Graph cut algorithm such as minimum spanning tree (MST), is then performed to assemble triangles into regions. At the last stage, the obtained regions are projected into the original image and region boundaries are refined by snake model. Experimental results show that edges are accurate, and a good spatial coherence is attained.