Kernel Discriminant Analysis (KDA) has been shown to be one of the promising approaches to handle the pose and illumination problem in face recognition. However, empirical results show that the performance for KDA on face recognition is sensitive to the kernel function and its parameters. Instead of following existing KDA methods in selecting popular kernel function, this paper proposes a new approach for constructing kernel using wavelet. By virtue of cubic B spline function, wavelet kernel function is constructed. A wavelet kernel based subspace linear discriminant (WKSLDA) algorithm is then developed for face recognition. Two human face databases, namely FERET and CMU PIE databases, are selected for evaluation. The results are encouraging. Comparing with the existing state-of-the-art RBF kernel based LDA methods, the proposed method gives superior results.