Abstract
Changes in the illumination condition cause dramatic variation in face appearance and seriously affect the performance of face recognition systems. This problem is addressed in the thesis by introducing an approach based on active Near-Infrared differential imaging. Assuming a static scene, a linear response of the sensor to the scene radiation and no saturation, it is shown theoretically and empirically in this thesis that the active differential imaging technique yields a face image independent of the change in ambient illumination. By taking the difference of two face images, one captured with the active illumination on and one with it off, the resulting image contains the face illuminated only by the active illumination source. This technique is compared with several representative illumination invariant face recognition techniques on a database containing faces captured under different illuminations and at different time. The results in face identification and verification experiments demonstrate the significant advantage of this Near-Infrared differential imaging technique over the other techniques. This thesis also presents a multistage approach to automatic face localisation for the Near-Infrared face images. This multistage approach is a combination of a novel pupil detection approach based on edge following and chaincode representation, and an approach based on FloatBoost learning. Accurate face localisation results are achieved by the proposed multistage approach, and this leads to the excellent face recognition performance in fully automatic scenario. A subject appears at two different locations before and after the active illumination is turned on if he/she is moving. This causes motion artifact in the difference image from the active differential imaging system and degrades the performance of the face recognition system. The thesis presents an approach based on motion compensation to deal with this problem. It is shown from the experimental results that the proposed approach successfully removes the motion artifacts and improves the face recognition performance significantly.