Robust 3D face recognition from low resolution images

This paper proposes a combined approach for robust face recognition from low resolution images captured by a low-budget commercial depth camera. The low resolution of the facial region of interest is compensated via oversampling techniques and efficient trimming algorithms for the generation of an accurate 3D facial model. Two state of the art algorithms for geometric feature extraction are then utilized, i.e. the estimation of the Directional Indices between all the isogeodasic stripes of the same facial surface via the 3D WeightedWalkthroughs (3DWW) transformation and the estimation of the Spherical Face Representation (SFR). The biometric signature is then enhanced via user-specific cohort biometric templates for each feature, respectively. The experiments have been carried out on the demanding "BIOTAFTOTITA" dataset and the results are very promising even under difficult scenarios (e.g. looking away instances, grimace, etc.). Despite the obvious superiority of the 3DWW transformation over the SFR, it has been noted that the score level fusion of both algorithms improves the authentication performance of the system. On the contrary, only the 3DWW transformation should be preferred in identification scenarios. Indicatively, the experimental validation on the aforementioned dataset containing 54 subjects illustrates significant succeeds an identification performance of ~ 100% in Rank-1 and Equal Error Rate of 0:25% regarding the authentication performance in the neutral face experiment.