Kernel uncorrelated neighborhood discriminative embedding for feature extraction
Xuelian Yu, Xuegang Wang · Optical Engineering · 2007
Feature extraction is a crucial step for pattern recognition. Recently, some manifold learning algorithms have drawn much attention. Although their properties of locality preserving are fairly significant, most manifold-based algorithms have limits to solve classification problems. First, they do not have good discriminant ability. Second, they fail to remove the redundancy among the extracted features. We present a new feature extraction method, called kernel uncorrelated neighborhood discriminative embedding (KUNDE), which integrates two abilities of manifold learning and pattern classification. The purpose of KUNDE is to preserve the within-class neighboring geometry while maximizing the between-class scatter. Optimizing an objective function in a kernel feature space, nonlinear features are extracted. Moreover, by putting a simple uncorrelated constraint on the computation of the basis vectors, the extracted features via KUNDE are statistically uncorrelated and thus contain minimum redundancy. Experimental results on radar target recognition indicate the promising performance of the proposed method.