利用特征空间主分量分解实现自适应小目标检测算法

The Adaptive Small Target Detection Based on Principal Component Analyzation in Feature Space

为了实现非平稳复杂背景干扰下的红外小目标检测,在特征选择阶段同时考虑了目标的强度、分布梯度以及与背景之间的阶越特性,并提取对比度、临域标准差以及多孔小波变换高频分量的模值三维特征予以描述.将目标检测过程视为特征空间中的两类分类问题,通过主分量分解获取背景的统计聚类参数,将目标像素特征向量视为背景特征聚类之外的异常点,并建立了统计距离作为检测判据.同时,以背景特征向量的多元正态统计分布为假设前提,利用各主分量偏度作为限制进行了用于小目标检测的统计距离阈值的选择,进一步提高了算法的自适应性能.实测试验证明了算法的有效性.

To detect infrared small target in non-stationary complex background,at first,the intensity of target,distribution gradient of target and step characteristic between target and background are considered,in process of feature selection.The three-dimensional features are extracted,which contain contrast between each local area pixel and its surrounding background,the local standard deviation of pixel and the modulus of high frequency components of wavelet transform at position of pixel.Regarding the process of target detection as binary classification problems in feature space,cluster parameter of background is acquired by principal component analyzation（PCA）,the feature vector of target pixel is treated as the abnormal point different from background cluster,and the statistical distance is built up as the criterion for detecting.The validity of this algorithm is proved by actual experiments.