This paper discusses facial recognition as applied to the classification of two-dimensional images and proposes a new architecture that allows a fast derivation of compact and invariant features from the Radon space. Our approach has been inspired by the review of feature-based non-connectionist and connectionist models of facial recognition. In the feature based non-connectionist model, a large part of the computational effort is focused on the extraction of facial features or the geometrical encoding of the face and the measurement of statistical parameters to describe their relationship. The connectionist model focuses on two-dimensional intensity values of the facial image allowing the geometrical encoding to be measured implicitly. The connectionist model is thus susceptible to variations in lighting conditions, spatial position and orientation of the images and can result in a poor detection of faces. An additional bottleneck of the connectionist model is the large feature vector size applied to its input that can cause non-convergence problems during training. The Radon transform is a generic transformation that is capable of representing shapes and it is used to compute harmonic components from which compact and invariant features can be derived. It is shown in this paper that these features when applied to a connectionist model result in a system that is capable of achieving high recognition rates and at high significance levels.