基于虚拟背景光谱Hausdorff距离的高光谱异常检测及GPU实现  

Hausdorff Distance with Virtual Background Spectrum for Hyperspectral Anomaly Detection and GPU Implementation

作　　者：赵春晖[1] 李佳伟 闫奕名 宿南 Zhao Chunhui;Li Jiazeaei;Yah Yinming;Su Nan(Information and Communication Engineering College,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001

出　　处：《沈阳大学学报（自然科学版）》2018年第4期284-293,共10页Journal of Shenyang University：Natural Science

基　　金：国家自然科学基金资助项目(61571145);黑龙江省自然科学基金重点资助项目(ZD201216)

摘　　要：提出了一种基于虚拟背景光谱Hausdorff距离的高光谱图像异常目标检测算法(VBS-HD).利用改进的部分Hausdorff测度计算待测像元与背景信息间各波段光谱的正反双向距离,并分别进行升序排列.为了避免数据成像过程中,由传感器、环境等因素产生的孤立噪声点对测试像元造成误判,算法选取前k个距离进行均值化,进行后续Hausdorff的度量.对于背景信息的选取,本算法根据待检测像元周围背景信息的特点,构造理想化虚拟背景光谱信息,从而有效减少背景信息掺杂噪声及异常点对检测性能的影响.为了进一步提高检测效率,在GPU/CUDA模型下,设计算法的并行系统架构,利用GPU的并行计算能力,获得较好的加速效果.通过模拟及真实数据的实验结果表明,VBS-HD异常检测算法具有较强的抗噪声干扰能力,在保证较高的检测精度的同时,显著提升了算法的执行效率.A new Hausdorff distance algorithm with anomaly detection is proposed. The improved partial virtual background spectrum for hyperspectral Hausdorff measure is used to calculate two way distance for positive and negative between pixel under test and background information each band spectrum and arranged in ascending order, respectively. In order to avoid misjudgment result from isolated noise produced by the sensor or environmental factor in data imaging process, the algorithm selects the first k distance to proceed equalization, then achieves the Hausdorff metric process. For the selection of background information, the algorithm according to the characteristics of the background information around pixel under test, structured idealized information of virtual background spectrum, thereby reduced the influence of background information mixed noise and abnormal points on detection performance effectively. In order to further improve the efficiency of detection, the parallel system architecture is designed based on GPU/CUDA model. It makes full use of GPU parallel computing capacity to accelerate the VBS HD algorithm. The experimental results with synthetic and real datasets show that the VBS HD anomaly detection algorithm has stronger ability to resist noise, and it can improve the accuracy as well as the efficiency.

关 键 词：高光谱图像处理 异常目标检测 HAUSDORFF距离 图像处理单元(GPU) 并行处理 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]