Unsupervised segmentation with CUDA for SAR imagery based on loop belief propagation

A novel segmentation approach for SAR imagery based loop belief propagation on Graphic Processor Unit (GPU) is presented in this paper. The approach combines Markov random field (MRF) and loop belief propagation (LBP) optimization. Our approach has two advantages. One is that parameters of distribution of different regions can be iteratively inferred without parameter estimation. Another is that our LBP is executed in parallel on GPU using the general-purpose computation ability of GPU, and the running speed of GPU-segmentation can be more than 80 times as fast as that of CPU-segmentation with the same performance. The experimental results show that the proposed method can improve computational efficiency greatly and provide precise segmentation results.