Face recognition using multi-scale differential invariants in statistical manifold framework

The local image structure can be robustly represented by multi-scale Gaussian derivatives (GDs) or the derived differential features. However, the high-dimensional nature of concatenated global features makes it hard to be applied directly. To utilise multi-scale Gaussian derivative-based differential invariants (MGDDI) up to order two for face recognition, a novel method of matching probabilistic generating model of MGDDI is developed in statistical manifold framework. It takes MGDDI of an image as multi-channel feature sets in which each one is univariate consisting of fixed dimensional components of local 'jets'. Under specific partitions on feature spaces, each channel feature set is modelled as a realisation of a marginal multinomial distribution, and corresponding normalised histogram can be identified with estimated model parameters. With the Fisher geometry on multinomial manifold, a similarity measure is proposed for matching marginal model sets. The effectiveness of proposed method is demonstrated by the promising experimental results on ORL and FERET face database.