一种改进的基于ICA特征空间的高空间分辨率影像的目标识别

Improved Recognition of High Spatial Resolution Images Based on ICA Feature Space

为解决高空间分辨率影像目标的识别问题,一种好的方式是将充分考虑高阶累积量的独立分量分析方法引入高空间分辨率影像进行特征提取,但由于基于传统独立成分分析方法提取的特征空间不能最优区分不同类别的样本。为此,提出一种改进的基于独立成分分析的目标识别方法(Multi-ICA)。该方法为每个类别的样本构造单独的特征空间,通过投影到特征空间,得到表征该类别样本特征的特征向量集合。Multi-ICA方法提取的特征空间是基于某类样本图像的共性特征建立的,同一类别样本间的欧式距离要小于不同类别样本之间的欧式距离。因此,可以将待识别样本分类到具有最小欧式距离的特征空间所对应的类别上。现以北京地区的高分辨率卫星Quickbird影像为例,进行了Multi-ICA、传统ICA方法、主成分分析(PCA)方法,以及Multi-PCA方法的目标识别对比实验。结果表明,提出Multi-ICA算法的识别率有明显的提高,并且在一定程度上缓解了由于样本数量增加导致样本特征向量维数增加的问题。

To solve the problems of high spatial resolution images＇ recognition, independent component analysis which takes full account of higher-order cumulants is introduced to extract the feature of high spatial resolution image. However, the feature space, extracted by traditional method based on independent component analysis, cannot optimally distinguish between different types of samples. Therefore, an improved algorithm based on independent component analysis （namely Multi-ICA） is proposed. In this algorithm, its own feature space for each type of sample is constructed, and then by projecting to the fea- ture spaces, the set of characteristic vectors representing the features of the given sample are obtained. The feature spaces extracted by the Multi-ICA algorithm are constructed according to commonness between the sample images of the certain type. The Euclidean distance between the same type of samples is smaller than that of the different type of samples. Taking the high-resolution Quickbird satellite image of Beijing district as a example, the contrast experiment of target identification shows that the recognition rate of the proposed Multi ICA algorithm in comparison with those of traditional ICA, PCA and Multi- PCA is more obviously improved. The recognition rate keeps stable when recognition types increases. It alleviates the problem that the dimension of sample feature vector increases with the increase of samples.