| Summary: | Human face as a biometric trait provides a natural, direct, acceptable and nonintrusive means of human identification. Face detection is an important issue for security and surveillance, access control, and human-computer interactions. Various techniques have been investigated aiming at high accuracy, reduced complexity and less computational time. Yet, occlusion continue to be one of the most challenging problem in face detection and recognition. These issues arise in real life situation because an individual face is not always neutral. In this thesis, an improved model capable of detecting face and facial features under various occlusions is proposed. The face region is extracted from the background using color segmentation. And Active Shape Model (ASM) is employed for face shape alignment, shape variation modeling and fitting the model on the salient features of the face. University Milano Bicocca (UMB) database is utilized for the experiment. This is an uncontrolled database with images of different poses, expressions, and illumination conditions. Most of the subjects have been with eyeglasses, holding phones, hat, partially occluded by hair, hand, scarf, and other miscellaneous objects. The major improvement of this study is centered on determining the proper range and color space of skin and also estimating the face occlusion based on the selected landmark points and their corresponding regions. In terms of speed and accuracy, there is an increase in rotation angle of the face by ± 15°, that makes the system more capable of detecting faces with different poses. Also, 25 iterations were considered which enabled the model to fit more accurately on the face. Images were annotated with 66 control points during training to represent the facial features accurately. The experiment results have indicated that the proposed model is effective in detecting face and facial features with different type of occlusions. The average fitting time, detection time and detection rate obtained are up to 0.42sec, 3.llsec, and 97.21% respectively. These results are more promising when compared to several state of the art techniques.
|
|---|