Semantic Segmentation for Remote Sensing Image Using the Multigranularity Object-Based Markov Random Field With Blinking Coefficient
Hongtai Yao, Le Zhao, Meng Tian, Yong Jin, Zhentao Hu, Qinglan Peng, Qian Qiu
Abstract
Semantic segmentation is one of the most important tasks in remote sensing. In the semantic segmentation of remote sensing images, some regions are repeatedly transformed between multi-classes, which affects the convergence speed and segmentation accuracy. This is because the increased spatial resolution makes the spectral distribution of geographic targets differ from the overall category distribution. Markov random field (MRF) model is widely used for semantic segmentation of remote sensing images because of its outstanding spatial description ability. Some scholars have made improvements on MRF models to extract more information or enhance semantic inference. However, these improvements fail to capture the correlation between the multi-granularity layers and the historical information. In this article, we propose a new MRF-based model, which adopts multi-granularity layers to realize the multi-granularity correlation representation of targets and the spatial-temporal inference of segmentation labels. First, the algorithm constructs a multi-grained layer structure based on remote sensing images to enhance feature extraction for targets of different sizes in images; secondly, for the multi-layer feature field, a cross-layer Gauss-Markov model is constructed based on intra-inter-layer feature correlation constraints; then, for the multi-granularity layer label field, a self-renewing pairwise spatial-temporal potential function with blinking coefficients is constructed based on the newly defined cross-layer augmented neighborhood system, which can accelerate the convergence of segmentation by using the history information and spatial neighborhood information. The proposed method is tested on texture images, SPOT-5, and Gaofen-2 images. Experiments show that the proposed method has better performance compared to other state-of-the-art MRF-based methods.