Block-based Deep Belief Networks for Face Recognition
Faculty of Computing, Health and Science
School of Engineering / Centre for Communications Engineering Research
This paper presents research findings on the use of Deep Belief Networks (DBNs) for face recognition. Experiments were conducted to compare the performance of a DBN trained using whole images with that of several DBN trained using image blocks. Image blocks are obtained when the face images are divided into smaller blocks. The objective of using image blocks is to improve the performance of the present DBN to visual variations. To test this hypothesis, the proposed block-based DBN was tested on different databases, which contain a variety of visual variations. Simulation results on these databases show that the proposed block-based DBN is effective against lighting variation. The proposed approach is also compared with other illumination invariant methods and was found to demonstrate higher recognition accuracies.