The human brain exhibits robustness against natural variability occurring in face images, yet the commonly attempted algorithms for face recognition are not modular and do not apply the principle of binary decisions made by the firing of neurons. This thesis presents a memory based face recognition method based on the concepts of local binary decisions and spatial change features. Local binary decisions are inspired from the binary conversions done by firing of neurons while spatial change features are inspired from the retinal processing of the human visual system. Applying these principles and by using the principle of modularity in a hierarchical manner, a class of memory based face recognition algorithms is formed. These algorithms when applied to difficult testing conditions show high recognition performance. This high recognition performance is enabled by (1) local binary decisions and (2) spatial change detection. The baseline algorithm formed by using these two concepts is called local binary decisions on similarity (LBDS) algorithm. An analysis is performed using the LBDS algorithm to optimize the parameters, and to study the relative effect of spatial change features, local binary decisions, normalization of features, normalization of similarity measure, use of color, localization error compensation and resolution on recognition performance. From the insights gained through the analysis, the LBDS algorithm is further improved by incorporating various preprocessing spatial filter operations to extract more spatial information. The inclusion of preprocessing step helps to achieve even higher recognition performance and robustness to difficult tasks. This improved algorithm is called enhanced local binary decisions on similarity (ELBDS) algorithm. The ELBDS algorithm is further used to incorporate the multiple training images per person in the gallery, and is called an exemplar based face recognition method. The following is the overall recognition performance when using single gallery image per person: 97% on AR, 100% on YALE, 97% on EYALE , 97% on CALTECH, 98% on FERET(FaFb), 94% on FERET(FaFc), 74% on FERET (FaDup1) and 76% on FERET(FaDup2). When using multiple training samples per person, following recognition accuracies are achieved, 99.0% on AR, 99.5% on FERET, 99.5% on ORL, 99.3% on EYALE, 100.0% on YALE and 100.0% on CALTECH face databases.