One of the most recent techniques developed for face recognition is diagonal principal component analysis (DiaPCA). In contrast to two-dimensional PCA (2DPCA), DiaPCA is based on diagonal representation of face images. Therefore, the correlations between variations of rows and those of columns can be kept. This paper proposes a novel projection scheme called bilateral DiaPCA (BDiaPCA) which simultaneously performs left and right projections of original face images in a new subspace derived from diagonal representation. Therefore BDiaPCA reduces the dimension of the original image matrix in both column and row directions. Experiments were carried out under varying illumination and facial expression using three well-known face databases: Yale, PF01, and a subset of FERET containing 200 people. Comparisons with 2DPCA and DiaPCA show that BDiaPCA outperforms these methods and achieves higher recognition rates using the Yang, Frobenius, or assembled matrix distance metrics.