Recognizing Faces by Dynamic Link Matching

Dynamic Link Matching is a neural dynamics for translation invariant object recognition that is robust against distortion. We here demonstrate human face recognition against a gallery of 112 neutral frontal view faces. Probe images are distorted due to rotation in depth and changing facial expression. Probe images and gallery models are represented by layers of neurons interpreted as labeled graphs. Nodes are labeled by local features based on the Gabor wavelet transform. Probe images are matched to the gallery of face models by Dynamic Link Matching. Concurrently with the matching process a simple winner-take-all mechanism identiies the correct model. A dynamic window of attention restricts the match to the part of the image occupied by the face.