Improving Single Sample Per Person Face Recognition With Augmentation and Probabilistic Image Enhancement Technique / Muhammad Tariq Siddique

Thesis submitted for the degree of Doctor In Philosophy Universiti Teknologi Brunei

ABSTRACT
As much as it is a challenge to use face recognition on Single Sample Per Person (SSPP), it is even a greater challenge when the single sample based face recognition is performed in an unconstrained environment. The unconstrained environment is normally considered as irregularities in facial expressions, pose, occlusion and illumination. This degree of difficulty increases due to the single sample and in the presence of occlusion and illumination. Extensive research has been done on face recognition under pose and expression changes. Comparatively, less research was reported on the occlusion and
illumination problems that occur in facial images. Occlusion may alter the appearance of face images and may cause deterioration in recognition. A robust method is required to handle the occlusion in the face image to improve the recognition performance. On the other hand, the illumination effects caused by poor lighting conditions in an unconstrained environment are known to considerably depreciate the performance of
otherwise capable face recognition systems. Thus, a robust pre-processing method is needed to deal with poor lighting and enhance the quality of images for face recognition.

The purpose of the proposed study is to implement an effective augmentation technique and image enhancement technique to enhance the performance of the SSPP face
recognition system in unconstrained environments. The proposed study first analyses the effect of classical augmentation techniques such as different geometric and photometric transformation methods. The virtual samples are created to expand the sample size to address the single sample problem. Then a local region-based technique is proposed to
deal with occlusion by creating virtual samples. To address the illumination in face images a probabilistic image enhancement has been adopted as a pre-processing approach. A deep neural network based FaceNet model is used to extract the features and support vector machine is used for classification. The performance of the proposed approaches is evaluated and exhibits superiority while handling occlusion and illumination over state-of-the-art counterparts.

Thesis (Degree) - Universiti Teknologi Brunei, 2024