基于模糊C均值聚类的大尺寸图像目标检测加速方法

Acceleration of Fuzzy C-Means Clustering Based Target Detection for Large Size Image

利用模糊C均值(FCM)聚类算法对大尺寸图像进行目标检测时,由于样本数量巨大,算法运行时间过长,不利于信息的及时处理。为提高大尺寸图像检测效率,给出了一个CPU+GPU平台下的详细加速方案。该方案利用CUDA并行技术,将FCM聚类等操作放在GPU端处理。同时,对只能在CPU端执行的操作,利用OpenMP技术并行。对四幅大尺寸(15884×3171)全极化SAR图像进行检测,平均加速约84.02倍。此外还利用MPI并行技术在双节点上实现了对四幅全极化图像的同时检测。

Fuzzy c-mean clustering method （FCM） is an unsupervised clustering algorithm, and its clustering process does not require any manual intervention. It has a certain advantage in images with uncertainty and fuzziness, and has been paid more and more attention in the field of target detection. However, with the increase of the size of the image, the sample set also increases dramatically, which leads to long computing time. Using CPU＋GPU platform to accelerate the FCM clustering algorithm is an effective method to shorten the clustering time. A detailed acceleration scheme on CPU＋GPU plat form is given, in which the FCM clustering and other operations are arranged on GPU side, and the oth ers are arranged on CPU side. CUDA and openMP parallel technologies are used to improve computation efficiency. In numerical experiments, we make use of 4 large full polar SAR images （15884×3171）, the average detection time is shortened from 287.44 seconds to 3.37 seconds, and the average speedup is 84. 02 on a single CPU＋GPU node. And we also use MPI technology to detect these images in parallel on two CPU＋GPU nodes. Experimental results show that under the CPU＋GPU platform, our accelera- tion scheme can speed up the target detection time greatly. It provides an effective solution for FCM based target detection for large size images.