A Top-Down Application of Multi-Resolution Markov Random Fields with Bilateral Information in Semantic Segmentation of Remote Sensing Images

This paper presents a new multi-resolution Markov Random Field (MRF) method for semantic segmentation of remote sensing images. The main contribution of this paper is to propose a new method of information interaction between the scales so that macroscopic information and microscopic information can be captured on each scale. First, we established a multi-scale structure. Second, in the modeling process of label field in each scale, we not only consider the spatial information between the pixels of the layer. But also take the spatial interaction between this layer and its upper and lower layers into account. Finally, using the most classic the maximum a posterior (MAP) criteria, start from the top level and solve it layer by layer. Experiments were performed on texture image, synthetic geographic image and remote sensing image. These experiments show that the proposed method provides a better performance than other Markov-based methods. (The accuracy increases by about 2%).