A method based on tree-structured Markov random field and a texture energy function for classification of remote sensing images

Abstract

A method for classification based on treestructured Markov random field (TS-MRF) and a texture energy function (TEF) is presented. The TEF consists of a second-order prior energy function with homogeneous-non internal and external fields obtained from 2-D Wold decomposition. Thus, in TEF is possible to characterize a combination of stochastic and structural texture. The TEF is used into binary MRF's defined in the TS-MRF model. The extended model TS-MRF/TEF is tested on remote sensing images for mangrove cover classification. An evaluation is carried out comparing results with a thematic map of study area. The classification results using TS-MRF/TEF are compared with results produced using supervised and unsupervised pixel-based classification methods. We show that incorporating the TEF into TS-MRF model the quality of the mangrove classification is improved.

Reference Fields Analysis of a Markov Random Field Model to Improve Image Segmentation

“…References [15,16] define a prior energy function by means of a non-homogeneous auto-model. The a priori energy function, or Gibbs energy, consists of a second-order energy function with clique potentials defined from texture fields obtained by means of the 2-D Wold decomposition [17].…”

“…We carried out experiments to analyze the influence of parameters such as internal and external reference fields in segmentation quality. Particularly, the performance of the Texture Energy Function (TEF) proposed in [15,16] is evaluated. The TEF uses an a priori energy function with non-homogeneous internal and where x s and x r are values at sites s and r for pixels in X, β sr is obtained by (1), v s єV and C 1 is the set of single-site cliques.…”

“…We can see that a slightly change of initial values of α and β may cause that final segmentation accuracy changes. In domains where segmentation depends on very similar object attributes, the problem complexity increases [15,23] and becomes desirable to increase segmentation accuracy in few pixels because on studies on land cover segmentation in remote sensing applications, the final amount of missegmented pixels involves km 2 of surface when they are corresponded to the ground. In these experiments, we can see from Figure 2 that using an a priori energy function with internal and external fields helps to obtain better accuracy that by using an energy function with only an internal reference field.…”

“…In this energy only pair-site cliques and a homogeneous internal reference field are used. In [15,16] a Texture Energy Function (TEF) for MRF models used for image segmentation is presented. The TEF consists of a likelihood model with Gaussian distribution p(y|x) and a second-order energy function with non-homogeneous internal and external fields obtained by means of the 2-D Wold decomposition [17].…”

Synthetic Aperture Radar Image Segmentation Based on Markov Random Field with Niche Genetic Algorithm

“…In recent years, Markov Random Field (MRF) is the most widely used image based statistical model in the applications of image edge detection, segmentation, as well as the resumption of texture analysis [1]- [2]. Image segmentation algorithm based on MRF is firstly introduced into image processing fields since S. Geman and D. Geman [3] proposed in 1984.…”