| Summary: | 碩士 === 國立中正大學 === 資訊工程研究所 === 91 === In this paper, a novel human perceptual watershed image segmentation method that is suit for different color spaces including grayscale images is proposed. This segmentation method is based on the nearest neighbor graph (NNG) and the drainage rainfall system. The proposed segmentation method has two stages. In the first stage, we transform an image to NNG by color dissimilarity between adjacent pixels, and then simulate the drainage rainfall system on NNG to partition the image into regions. The pixels in a region will be connected and similar to each other in human visual system. In the second stage, we integrated similar regions produced in the first stage into meaningful regions by the properties of NNG. During the first stage, a blurring-like process “perception dissimilarity thresholding” and a multiple NNG (MNNG) technique are used to decrease the amount of regions produced in this stage and preserve the boundaries between interested objects in the image. During the second stage, we introduce a progressively merging algorithm to enhance the integrated segmentation results and a new formula of color dissimilarity to reduce the impact of noises and to obtain textured objects in an image. Mostly, each segmented region in an image can represent an object in the real world.
|
|---|
