Visual filters for face recognition

The authors describe a general approach for the multiscale representation, detection, and recognition of object primitives as it applies to face recognition tasks. The approach is based on radially non-uniform sampling strategy, and a local light adaptation mechanism for low-level image representation. Early processing involves feature encoding and classification using visual filter banks implemented via self-organizing feature maps (SOFM). Optimal filters are constructed by means of an iterative, crossvalidation-like data reduction algorithm. The derived visual filter representation is applicable to both (i) facial landmark detection, and (ii) face identification. Experimental results on a data set of over 200 images prove the feasibility of their approach.