Parallel Unsupervised SAR Image Change Detection Based On Intel MIC

SAR(Synthetic Aperture Radar) image change detection is to detect the change of the image in the same area at different times. It has a wide range of applications in various fields. For example, the change of forest coverage, urban expansion, land use, the change of wetlands, deserts, swamps and other special lots, growth status of crop, etc. They can all be made effective evaluation by image change detection. Therefore, many scholars at home and abroad have done depth research on this subject. With the increasing scale of SAR image, traditional serial algorithm in dealing with SAR image change detection has been low efficiency. It is necessary to change the traditional serial algorithm to parallel algorithm in order to improve efficiency. In recent years, heterogeneous computing develops rapidly. Compared with multi-core processor, MIC architecture has more cores,wider vector units and hardware thread. So MIC architecture is capable of handling more complex parallel computing. For example, Intel Xeon Phi 7110 P has 61 cores and 244 threads. You can make this 244 threads deal with parallel computing tasks simultaneously.Xeon Phi 7110 P has great advantages in parallelization. In this paper, we present parallel unsupervised SAR image change detection based on Intel MIC. The main work is as follows:(1)We introduce traditional serial FLICM(Fuzzy Local Information C-Means)algorithm.It can meet time efficiency when dealing with small-scale image. But with the increase of image size, the time efficiency is getting worse. To solve this problem, we use Open MP(Open Multi-Processing) framework on multi-core CPU and MIC to implement parallel tasks. First we use analysis tool to find the most time-consuming function or loop of the algorithm. In front, we add parallel control statement. Then we improve the internal structure of function or loop to achieve better parallel effect. This paper focus on accelerating the speed of parallel algorithm on multi-core CPU and MIC. For example, it has been found that after using Open MP to parallelize the loop, the outermost layer and second outer layer is not data-dependent. So we merger the circulation to achieve optimization; frequently used function has been rewritten to achieve better acceleration effect. Experimental results show,in Intel Xeon Phi, the image with the size of 2048 ×2048 can obtain 10.39 times speedup.(2)We introduce FGKM(Fast Global K-Means) algorithm. It’s inefficient when dealing with large SAR image change detection. To solve this problem, this paper presents parallel unsupervised SAR image change detection based on FGKM algorithm. We take advantage of multi-thread of MIC architecture. The large-scale SAR images are assigned to each thread to achieve parallel task. We first propose parallel SAR image change detection based on multi-core CPU and use Open MP framework to implement parallel tasks. Then we propose parallel SAR image change detection based on MIC-offload. We run serial algorithms with lower degree of complexity on CPU and transplant more time-consuming part into MIC to achieve parallel tasks. We combine CPU with MIC to accelerate our algorithm. Finally, we make optimization on the parallel algorithms on multi-core CPU and MIC-offload to achieve higher speedup. Experimental results show,in Intel Xeon Phi, the image with the size of 512 × 512 can obtain 22.47 times speedup.