Texture based automatic face recognition (AFR) methods proposed in the last few years have been successful in large-scale applications where the database consists of a single frontal view per person. In those methods the global similarity between two faces is generally given by a linear combination of the local similarities computed upon each face region. Little attention has been given so far to the estimation of the weights that express the relative contribution of each face region to global similarity score. This paper addresses this issue and proposes a method to estimate the optimum weighting for texture based AFR. The solution is given by the most discriminative axis within a similarity space using Fisher discriminant analysis. The proposed method is evaluated in experiments conducted on the FERET and on the FEI face databases. For texture coding both Local Binary Patterns (LBP) and Local Phase Quantization (LPQ) are considered. The experiments indicate that the proposed method brings a substantial improvement in terms of recognition performance in comparison to other weightings and weighting methods proposed in the literature.