Research On The Parallelization Of Mean Shift Remote Sensing Image Segmentation Algorithm Based On GPU Cluster

In last decades, a variety of remote sensing image segmentation algorithm is been proposed, but there are still low accuracy, poor adaptability problems exiting in remote sensing image segmentation segmentation. Mean shift algorithm is a characteristicvector-based clustering algorithm, and has been widely used in such fields like target tracking, image noise smoothing, image segmentation etc. Although the Mean Shift image segmentation algorithm is adaptable, and it has good segmentation accuracy, the processing of the mean-shift-based image segmentation algorithm is compute-intensive because its required computation grows exponentially with the involved amount of image pixels. In order to enhance the efficiency of the mean shift algorithm, many researchers have devoted their efforts on improving the processing efficiency by designing the corresponding parallel algorithm on single computing node that equipped with one or multiple graphics processing unit(GPU) cards with CUDA(compute unified device architecture). However, these researches face the following two challenges. Firstly, the parallel algorithm utilizing CUDA has poor portability and versatility. That is, such parallel algorithms can only run on the products of NVIDIA Corporation rather than on GPU cards made by other vendors, such as AMD and Intel. Secondly, the computing tasks demand more than one node that contains multiple GPU cards in some specific massively remote sensing image processing applications.To resolve the performance bottleneck of the mean-shift-based image segmentation algorithm, in this paper, we design and implement a parallel algorithm suitable for a heterogeneous GPUs cluster platform with more than one GPU-accelerated computing node. On such a heterogeneous computing system, all the computing units including both CPUs and GPUs shall work together collaboratively to maximize the overall performance. The main contributions of this study are as follows.(1) Design and implement a parallel mean-shift-based image segmentation algorithm, and explore its performance optimization methods on one single GPU node. This procedure consists of several steps. Step 1, implement the serial algorithm that can run on Linux environment; Step 2, locate the hotspot of the serial algorithm by using performance analysis tools, such as Intel? Vtune Performance Analyzer. Step 3, analyze the algorithm and its performance bottleneck, and chose the suitable parallelism strategies. Step 4, outline the parallel framework and flowchart, and implement the corresponding parallel algorithm with OpenCL, a crossing-platform programming model for heterogeneous computing. And step 5, further optimize the performance of the parallel algorithm from the aspects of data transmission and distribution.(2) Migrate the parallel algorithm to the heterogeneous GPUs cluster platform and propose a suitable task scheduling/load balancing strategy for massively remote sensing image processing in applications. This strategy is the combination of two programming models, MPI(Message Passing Interface) and OpenCL. Among different computing nodes, MPI is used for coarse-grained task segmentation. On one GPU computing node, OpenCL is utilized for fine-grained load balancing and calculation among different computing units.(3) Validate the correctness and effectiveness of the proposed methodology with a specific application. In this application, we need to detect the temporal remote sensing image change with several remote sensing images.From the testing results, it shows that:(1) On single GPU platform, the designed parallel algorithm can achieve a speedup of 30 X.(2) The proposed optimization algorithm resolved the correlation issue of the obtained speedup with the number of OpenCL work items so that the achieved speedup will fall into one certain range stably.(3) On the heterogeneous GPU cluster, the obtained speedup increases lineally with the number of the GPU nodes. This demonstrates that our parallelization strategy works effectively.(4) The accuracy achieved in applying our algorithm to a real world application demonstrates that our methodology is feasible for those change detection applications which need mean-shift-based image segmentation algorithm to process multi-temporal remote sensing images.