Distortion invariant object recognition by matching hierarchically labeled graphs

A graph-matching process of object recognition is proposed. It is applied to face recognition. Gray-level images are represented by a resolution hierarchy of local Gabor components, which are all scaled and rotated versions of each other. The components centered on one image point form a Gabor jet. A single jet provides a distortion-insensitive local representation of part of an image. Object recognition is achieved by matching image point jets to jets in stored prototype patterns. For a selected image jet the best matches are determined, under a constraint preserving spatial arrangement. The procedure amounts to labeled graph matching, with Gabor jets forming labels to nodes and topology determining links. A contrast-insensitive similarity measure provides for invariance with respect to lighting conditions. The authors have formulated the matching procedure as an optimization task solved by diffusion of match points. This diffusion is controlled by a potential determined by jet similarity and the topology-preserving constraint. The algorithm implements a neural network architecture.<>