Maximum Margin Criterion Embedded Partial Least Square Regression for Linear and Nonlinear Discrimination

More recently, the partial least square regression (PLSR) has been suggested applying to pattern discrimination. However, the eigen-structure problem essential to the discriminant PLSR basically depends on a slightly modified version of the between-class scatter matrix S_b. Unfortunately, the class structure information contained in the within-class matrix S_w is skipped when using PLSR for discrimination. To overcome this drawback, this paper presents a new scheme for pattern classification by incorporating the maximum margin criterion (MMC) into the PLSR (refered to as PLSR/MMC). We further extend the PLSR/MMC to its nonlinear domain via the kernel trick. The scheme given in this paper essentially describe an approach wherein the various advantages of the MMC and PLSR are combined to augment each other. The experiments on both face recognition and facial expression recognition have shown the superiority of the proposed method over the conventional PLSR