A label embedding kernel method for multi-view canonical correlation analysis

In this paper, we propose a novel label embedding kernel method (LEKM), which is capable of well capturing intrinsic discriminating structure of samples with the help of class label information. LEKM can efficiently project training samples into a label kernel space according to a label-based unit hypersphere model. However, it is difficult for LEKM to map out-of-sample data into the label kernel space due to the lack of out-of-sample class label information. To solve the problem, we give a simple but effective fuzzy projection strategy (FPS) that can approximately project out-of-sample data into the label kernel space according to similarity principle of sample distribution. With LEKM and FPS, we present a label embedding kernel multi-view canonical correlation analysis (LEKMCCA) algorithm, which can extract nonlinear canonical features with well discriminating power. The algorithm is applied to object, face and handwritten image recognition. Extensive experiments on several real-world image datasets have demonstrated the superior performance of the algorithm.