| Summary: | 碩士 === 國立臺灣大學 === 電信工程學研究所 === 104 === In computer vision, image segmentation plays an important role due to its widespread applications such as object tracking and image compression. Image segmentation is a process of clustering pixels into homogeneous and salient regions, and a number of image segmentation algorithms and techniques have been developed for different applications. To segment an image accurately with the number of regions user gives, we propose an adaptive growing and merging algorithm. Our procedure is described as follows: First, a superpixel segmentation is applied to the original image to reduce the computation time and provide helpful regional information. Second, we exploit the color histogram and textures to measure the similarity between two adjacent superpixels. Then we conduct the superpixel growing based on the similarity under the constraint of the edge’s intensity. Finally, we generate a dissimilarity matrix for the entire image according to color, texture, contours, saliency values and region size, and subsequently merge regions in the order of the dissimilarity. The region merging process is adaptive to the number of regions and local image features.
After the superpixel growing has been finished, some superpixels expand to larger regions, which contain more accurate edges and regional information such as mean color and texture, to help with the final process of region merging. Simulations show that our proposed method segments most of images well and outperforms state-of-the-art methods.
|
|---|
