The Research On Image Segmentation Algorithms Based On RGBD Images

As an important research direction of image processing, image segmentation has been widely used in medical image processing, traffic image processing, medical image processing and other fields. At the same time, image segmentation is the basis of image analysis and understanding. So image segmentation is important both in research and application. Currently, the image segmentation research mainly focus on the color images, but color images fail to provide the 3D space information, thus have certain limitations. With the emergence of RGBD images, they can provide both the color and depth information of the scene, thus attract extensive concern and become a new hotspot in the field of image segmentation. This dissertation focuses on image segmentation algorithm based on RGBD images. For color images, PB algorithm is used to detect the edge information. For depth images, the edge detection operators for depth and normal are proposed. Based on these edge detection operators, a marked watershed algorithm is presented to segment the RGBD images. The algorithm can partition regions having similar color and texture and improve the accuracy of segmentation. The main work of this dissertation is listed as follows:(1) Image segmentation algorithms and existing RGBD datasets are concluded and analyzed:Several representative segmentation algorithms, including their segmentation principles and characteristics are expounded. Several existing RGBD datasets and their characteristics are introduced and analyzed.(2) The edge detection operators based on RGBD images are proposed:For color images, an edge detection algorithm of multiple gradients fusion is used to make full use of the different attributes of pixels. For depth images, starting from the 3D geometry information, two edge detection operators are proposed to capture the changes of depth information and normal information.(3) A marked watershed algorithm for RGBD images is proposed:In order to solve the over-segmentation in watershed segmentation algorithm, marker images are extracted by fusing three different gradient images and are used to modify the color gradient images. The algorithm can reserve different kinds of edges information in the images, and has good segmentation results on the regions having similar color and texture.