In this paper, we propose a new approach to recognize 2D faces. This approach is based on experiments performed in the field of cognitive science to understand how people recognize a face. To extract features, the image is first decomposed on a base of wavelets using four-level Difference Of Gaussians (DOGs) functions which are a good modeling of human visual system; then different Regions Of Interest (ROIs) are selected on each scale, related to the cognitive method we refer to. After that, Local Binary Patterns (LBP) histograms are computed on each block of the ROIs and concatenated to form the final feature vector. Matching is performed by means of a weighted distance. Weighting coefficients are chosen based on results of psychovisual experiments in which the task assigned to observers was to recognize people. Proposed approach was tested on IV2 database and experimental results prove its efficiency when compared to classical face recognition algorithms.