Face images always have significant intra-class variations due to different poses, illuminations and facial expressions. These variations trigger substantial deviation from the linearity assumption of data structure, which is essential in formulating linear dimension reduction technique. In this paper, we present a kernel based regularized graph embedding dimension reduction technique, known as kernel-based Regularized Neighbourhood Preserving Embedding (KRNPE) to address this problem. KRNPE first exploits kernel function to unfold the nonlinear intrinsic facial data structure. Neighbourhood Preserving Embedding, a graph embedding based linear dimension reduction technique, is then regulated based on Adaptive Locality Preserving Regulation Model, established in  to enhance the locality preserving capability of the projection features, leading to better discriminating capability and generalization performance. Experimental results on PIE and FERET face databases validate the effectiveness of KRNPE.