A Study of Face Alignment Methods in Unmasked and Masked Face Recognition

Abstract: Face recognition is a popular technology that has been widely used in various applications, such as security systems, social media, and entertainment. However, recognizing faces accurately is challenging due to factors such as facial expressions, head poses, and lighting conditions. Face alignment is a critical sub-process in the face recognition pipeline, which aims to normalize the facial image by correcting for variations caused by facial expressions and postures.  Although face alignment has been shown to improve the accuracy of face recognition, there are different methods available, and it is unclear how these methods affect the performance of face recognition approaches. Furthermore, the effectiveness of applying face alignment in other face recognition problems, such as masked face recognition and face clustering, remains uncertain.  To address these issues, this master thesis compares three commonly used 2D face alignment methods in various face recognition problems, including common face recognition, masked face recognition, and face clustering. Different deep face recognition approaches are also considered to evaluate the effectiveness of the alignment methods. The results show that affine transformation-based face alignment is the most effective method for common face recognition and clustering problems, regardless of the face recognition approach used. However, applying face alignment methods may reduce performance in masked-face recognition.