Alignment of Facial Images: Predicting Dense Displacement Fields for Temporal Facial Image Registration

Abstract

Clinical facial image analysis relies on accurate alignment of image pairs captured at different time points to monitor dermatosis changes. Although images are acquired under controlled conditions, misalignment still occurs due to temporal variations, occlusions, and orientation differences, posing a challenge for reliable facial skin analysis. To address this problem, this thesis investigated approaches for pixel level alignment in the clinical facial domain. This study demonstrates that WAFT, pre-trained on the Spring dataset, serves as a robust baseline for clinical facial alignment when initial alignment is within an acceptable range. Furthermore, two strategies to improve WAFT are explored: supervised fine-tuning on synthetic facial data and UWAFT, an unsupervised variant designed for real-world clinical datasets. While supervised fine-tuning generalizes well to synthetic data, it fails to adapt to real-world scenarios. The unsupervised fine-tuning faced limitations due to dataset size and quality, emphasizing the need for domain-specific data and loss design. Visual results indicate that WAFT can potentially be further improved through additional fine-tuning on clinical facial data. This study identifies WAFT, pre-trained on the Spring dataset, as an effective baseline and introduces UWAFT as a step toward further domain adaption using unlabeled clinical facial data. These findings provide insights into strategies for improving alignment accuracy and evaluation, contributing to the advancement of facial image registration techniques for dermatological applications and laying the foundation for future real-world implementations.