数据场典型相关分析及其在图像分割中的应用

针对数据场环境下多维数据的低维特征提取问题,本文将数据之间的相互作用纳入其相关性求解中,提出一种基于数据场的典型相关分析(Data field based canonical correlation analysis, DFCCA)方法. DFCCA提取的特征具有良好的分布特性,原空间上相隔较远的数据点对的特征聚集在一个较小区域内,而相邻数据点对的特征却有规律地分布在其他点所聚集区域的周围.此特性使得DFCCA具有较好的边界辨识能力,将其应用于图像分割的实验结果表明, DFCCA提取的复杂图像边界具有较好的保真度.

Data Field Based Canonical Correlation Analysis and Its Application to Image Segmentation

In this paper, for extracting low-dimensional features from multi-dimensional data in data field environment, we propose a novel method of canonical correlation analysis(CCA) called DFCCA(data field based CCA) by introducing interactions among data into data correlation solving. The features extracted by DFCCA have better distribution properties, that is the features corresponding to a data point pair that are far apart from each other gather together in a small region, but other features corresponding to the pair of data points that are neighboring each other will scatter regularly around the region. Thanks to these properties, DFCCA has a good capability of frontier identification. Experimental results on image segmentation demonstrate that the frontiers extracted from complex images by DFCCA hold better fidelity.