Object-Based and Semantic Image Segmentation Using MRF
- PDF / 4,469,351 Bytes
- 8 Pages / 600 x 792 pts Page_size
- 88 Downloads / 209 Views
Object-Based and Semantic Image Segmentation Using MRF Feng Li Shanghai Zhongke Mobile Communication Research Center, Shanghai Division, Institute of Computing Technology, Chinese Academy of Sciences, Shanghai 201203, China Institute for Pattern Recognition & Artificial Intelligence, State Education Commission Laboratory for Image Processing & Intelligence Control, Huazhong University of Science and Technology, Wuhan 430074, China Email: [email protected]
Jiaxiong Peng Institute for Pattern Recognition & Artificial Intelligence, State Education Commission Laboratory for Image Processing & Intelligence Control, Huazhong University of Science and Technology, Wuhan 430074, China Email: [email protected]
Xiaojun Zheng Shanghai Zhongke Mobile Communication Research Center, Shanghai Division, Institute of Computing Technology, Chinese Academy of Sciences, Shanghai 201203, China Email: [email protected] Received 6 December 2002; Revised 3 September 2003 The problem that the Markov random field (MRF) model captures the structural as well as the stochastic textures for remote sensing image segmentation is considered. As the one-point clique, namely, the external field, reflects the priori knowledge of the relative likelihood of the different region types which is often unknown, one would like to consider only two-pairwise clique in the texture. To this end, the MRF model cannot satisfactorily capture the structural component of the texture. In order to capture the structural texture, in this paper, a reference image is used as the external field. This reference image is obtained by Wold model decomposition which produces a purely random texture image and structural texture image from the original image. The structural component depicts the periodicity and directionality characteristics of the texture, while the former describes the stochastic. Furthermore, in order to achieve a good result of segmentation, such as improving smoothness of the texture edge, the proportion between the external and internal fields should be estimated by regarding it as a parameter of the MRF model. Due to periodicity of the structural texture, a useful by-product is that some long-range interaction is also taken into account. In addition, in order to reduce computation, a modified version of parameter estimation method is presented. Experimental results on remote sensing image demonstrating the performance of the algorithm are presented. Keywords and phrases: semantic and structural segmentation, MRF, Wold model, remote sensing image.
1.
INTRODUCTION
In this paper, remote sensing image segmentation based on the Markov random field (MRF) is considered. Many approaches have used MRF as a label process (as discussed in [1, 2, 3, 4, 5, 6, 7, 8, 9]), including the application to extract urban areas in remote sensing images (as discussed elsewhere in [5, 10, 11]). This is because exploiting MRF offers several advantages over simple segmentation algorithms. First, the segmentation for the object in a remote sensing image depends not only on the
Data Loading...
