Acceleration Of Images Segmentation Algorithm Using Non-Gaussian Triplet Markov Random Fields Models Based On GPU

Synthetic Aperture Rader(SAR) image segmentation is an important stage forSAR images’ recognition and understanding, and the research for the SAR imagesegmentation algorithm has been a hot spot. Since the SAR image contains a largenumber of multiplicative speckle noise， it is difficult to apply traditional imagesegmentation algorithm. Given the advantages of the Triplet Markov randomFields(TMF) models and the statistical properties of SAR images, TMF model is usedfor non-stationary, non-Gaussian images modeling.Non-Gaussian TMF image segmentation algorithm has achieved very good resultsin practical applications, but the complexity of the algorithm is high, operating speeddoes not lead to satisfactory. It is contrary to real-time and efficiency of SAR imageprocessing. In this paper, a parallel method based on GPU-HPC is presented, it willparallelized non-Gaussian TMF segmentation algorithm. Since restricted by thetraditional hardwired, GPU is not efficient in parallel computation. While CUDAreleased by NVIDIA Corporation has higher programming performance, widerapplication area and supports better CUDA hardware performance compared withtraditional GPU. Through parallelizing non-Gaussian TMF segmentation algorithmusing CUDA, efficient using CPU and GPU resources for parallel computing, throughGPU hardware acceleration, segmentation results in this paper show good performancewith high accuracy and in short execution time.Meanwhile, to improve the operating efficiency of the program, this paperevaluates the utilization of GPU. Analyzes the specific hardware structure, optimizesCUDA code and improve its structure, further enhances the performance of non-Gaussian TMF segmentation algorithm.