Robust Face Recognition

Abstract

The objective of this thesis is to propose new algorithms in face detection which have the capabilities of detecting the face with different poses and under different conditions. This objective is obtained in different stages and using different proposed algorithms. Firstly, a robust segmentation algorithm is proposed to segment the skin region from the image. Secondly, different filtering steps are applied to this segmented image to obtain the face candidate region only. After that, Feature-Based approach is used to detect the features from this candidate face which can work in real-time with minimal training in contrast to other approaches such as image-based approach. Finally, some rules is applied in order to judge if this candidate is profile face or not either the profile face is right or left. To strengthen this work, Template Matching approach is used to detect a frontal face. After detecting the face correctly, some pre-processing steps is applied to the detected face to obtain a down scaled normalized gray face image that help in reduce the computation complexity when applying the face recognition algorithm. The Fisher faces technique is applied to the processed face in order to recognize it in different conditions. This work has advantages over other approaches flow out from the ability of the proposed algorithm on detecting faces with different poses (Left, Right and Frontal) in varying illumination conditions and different races. This is done by merging some already defined algorithms such as Template Matching with new proposed algorithm to strength the detection process and obtaining robust algorithm with good results. Experimental results show that the proposed method is robust under a wide range of lighting conditions, different poses, and different races. These results are taken from three different face databases which are GTAV, FEI and Champions. These databases contain peoples from different races under different illumination conditions. The proposed method is implemented using Matlab version 7.6 software and they evaluated using Delll-inspiron laptop run with a 2-GHz CPU and has a 3-GB DDR2 RAM. The correct detection rate for the GTAV face database is 95.04%. While, for FEI face database it has 94.5% rate. Finally, For Champions face database the detection rate is 93.3%. The proposed face detection algorithm give good results compared to other researchers algorithms. The comparison of this proposed algorithm with other face detection algorithms gives an improved performance rate estimated by 2.87%. While, the comparison of our used face recognition algorithm after applying some proposed pre-processing steps with other algorithms gives an improved performance rate estimated by 0.2%.