基于带野值的SVDD的高光谱图像异常检测  

Anomaly Detection for Hyperspectral Image Based on SVDD with Negative Examples

作　　者：蒲晓丰[1,2] 雷武虎[1,2] 汤俊杰[3] 黄涛[1,2] 

机构地区：[1]电子工程学院脉冲功率激光技术国家重点实验室,合肥230037 [2]电子工程学院506教研室,合肥230037 [3]电子工程学院电子系,合肥230037

出　　处：《光电工程》2010年第12期83-87,共5页Opto-Electronic Engineering

基　　金：安徽省自然科学基金资助项目(070415217)

摘　　要：基于支持向量数据描述(SVDD)的高光谱图像异常检测算法常常会因为背景样本中混入异常像元(野值)而导致检测概率下降。针对此问题,提出一种基于带野值的支持向量描述的检测算法,引入原点和少量已检测出来的异常像元作为异常样本,以改善算法对异常和背景数据的描述能力;同时为抑制错误样本的影响,将异常和背景样本偏离背景样本均值的距离映射作为各样本的加权系数。结果表明,新算法在低虚警概率下较之SVDD模型有更高的检测概率。利用真实数据进行实验证明了算法的有效性。Anomaly detection algorithm based on the Support Vector Data Description （SVDD） usually brings low detection rates due to background examples being contaminated by anomalous data. To deal with the problem, a new method based on SVDD with negative examples is proposed. By leading into the origin and a few anomalies as negative examples, the capability of description to anomaly and background is improved. In order to control the influence of false examples, the distance of each sample deviating from the mean of background example is mapped into the sample as weighting coefficient. The results show that the proposed method could obtain higher detection rate under low false rate than the algorithm based on SVDD. The effectiveness of the proposed method is validated by experimental results obtained from real data.

关 键 词：异常检测 支持向量数据描述(SVDD) 加权 高光谱图像 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]