The segmentation-based image compression technique is expected to be one of the influential methods to overcome the defects of conventional block-based approach in low-bit rates. The main research items in this area are largely classified into two categories. One is to develop an efficient scene segmentation method, and the other is to provide coding tools to reduce region information. This dissertation is devoted to these topics including an image segmentation using a new similarity criterion and a relative shape approximation method for shape parameters coding.
First, an image segmentation algorithm which divides a scene into homogeneous local regions is investigated. Based on the observation that the conventional criterion considering motion or intensity homogeneity alone leads to partially incorrect results, the structure combining these two effects in a single segmentation procedure is presented. The actual segmentation is accomplished using the region growing technique based on watershed algorithm. It comprises processes of extracting the initial regions called seeds and their growing by merging similar neighbor pixels. Simulation results show that the proposed method is effective in determining object boundaries not easily found using static or dynamic criterion alone.
Secondly, a tracking technique is presented to make a coherent segmentation through time. Typical image sequences are greatly correlated between adjacent frames. Thus, coherent segmentation through frames is essential to achieve high compression rate by the elimination of the temporal redundancy. For this purpose, a tracking algorithm that comprises new object detection, temporal projection, and boundary readjustment processes is presented. It is examined that the proposed method solves the temporal coherency and region correspondence problems through some simulations.
Finally, an adaptive approximation method of contour signal generated by the region-based segmentation is presented. The method approximates contour with polygons using variable $d_{max}$ determined by contrast of regions. By the method, the high contrast part of contour is accurately represented while saving vertices in the low contrast part. Using this scheme, the texture error caused in the error region can be reduced by about 10-30% compared to the conventional one.
