基于投影寻踪的高光谱图像异常检测并行算法

Parallel algorithm of anomalies detection in hyperspectral image with projection pursuit

投影寻踪方法能有效提取数据中的非高斯结构凸显异常信息,但在求解最优投影方向时存在计算量大、运行时间长的问题,为提高处理效率,提出一种机群环境下的并行算法.选用偏度和峰度组合作为投影指标,将所有像素光谱作为特定投影方向集依次搜索,求解最优投影方向.在并行计算各候选方向投影指标时,分割图像数据分布存储于各机群结点,数据子块朝候选方向并行投影后,将指标计算式变形分解,使各结点在指标计算过程中所需数据均为本地数据,解决数据局部性问题,并采用一种"轮流作主"的机制提高算法负载均衡程度.利用实用型模块化高光谱仪数据在机群系统上进行测试,达到了较好的加速效果,表明该并行算法具有良好的并行性能.

Projection pursiut can extract non-Gaussian structure in hyperspectral data to reveal the anomalies information, but searching the best projection directions is a computational intensive task. To improve the process efficiency, a parallel process algorithm under cluster system was presented. The combination of skewness and kurtosis was selected as projection index （PI）. Using all the pixels＇ spectral as a special projection dierection set and searching the best projection directions in it. While parallel computing the PI value for each candidate dierection, the hyperspectral data was distributed to each computing node after partitioned evenly. After projecting each data subblock to a candidate direction in parallel, the index computation was transformed and decomposed. This makes all the data needed during index computation be in local memory for each node and decreases the communication. Furthermore, a ＂be host in turn＂ method was put forward to improve the degree of load balance. Using an operative modular imaging spectrometer data to test the efficiency on cluster, the results show that the parallel algorithm achieves good parallel performance.