A multi modal approach towards pose and illumination invariant face recognition from video still images

Abstract

newlineRecognizing the Faces from degraded video frames or still images is one of the most discussed research problem in computer vision. Face Recognition from clear images has reached maturity level using Deep Learning methods which works with very large Face Database. In a considerable lot of circumstances, it is exceptionally hard to get huge number of Face Images for the confirmation purpose, particularly from Village individuals. newlineThe development of Face Recognition systems is still restricted by the conditions achieved by several real applications even though current Face Recognition systems have attained a moderate level of maturity. The main bottle-necks in Face Recognition are Pose variation, Illumination variation, Occlusion, Noise and Blur. In the thesis we propose different methods to recognize the Face from degraded video /still images. The image may be degraded due to motion blur, noise or variation in illumination. Another important problem to be addressed is Pose variation. Face Recognition methods still needs more efficient technique for addressing Pose variation, although many research works tried to overcome this problem. In our thesis, we tried to address all the said problems by implementing necessary algorithms. We introduced six different Face Recognition approaches , 1) Local Linear Regression and Facial Recognition, 2) Blur Normalization and Facial Recognition ,3) Face Recognition with Support Vector Machine and Particle Swarm Optimization , 4) Hybrid Dense Matching Feature Based Video-Face Recognition with Modified Support Vector Machine approach,5) Face Recognition Across Noise, Blur And Illumination Using Deep Learning Neural Networks, 6) Blur and Noise Removal from the Degraded Face Images for Identifying the Faces using Deep Learning Networks newlineIn the first approach we address the illumination and pose problems for unconstrained Face Recognition task with Discrete Cosine Transform (DCT) and Local Linear Regression (LLR) respectively.