Kernelized discriminative canonical correlation analysis

Feature extraction using canonical correlation analysis (CCA) manipulates the pairwise samples from two information channels, say, X and Y, respectively, to realize the feature fusion in the context of multimodal recognition. To extract more discriminative features for recognition, a new supervised kernel-based learning method, namely kernelized discriminative CCA (KDCCA), is proposed. The superiority of KDCCA to CCA lies in 1) the class information is well exploited so that KDCCA is a supervised learning method; 2) the kernel method is employed to tackle the linearly inseparable problem in real applications. The experiments validate the effectiveness of KDCCA and its superiority to CCA and its kernel version in terms of the recognition performance.