Akademik Veri Yönetim Sistemi
GITMA, Sakarya, Türkiye, 2 - 04 Haziran 2025, ss.60-78, (Tam Metin Bildiri)
This research aims to investigate and address the challenges of data loss in single-image texture completion caused by head rotations, particularly vertical rotations. Using Shannon entropy as a metric, we plan to quantify the extent of information loss and evaluate the limitations of existing texture completion methods, such as OsTeC, in handling occlusions. We hypothesize that vertical head rotations result in greater data loss than horizontal rotations, leading to lower reconstruction accuracy. To mitigate these challenges, we propose leveraging Denoising Diffusion Probabilistic Models (DDPMs) to reconstruct missing textures and improve texture completion accuracy. Our methodology involves training a diffusion-based neural network using the CASIA-WebFace dataset and comparing its performance to OsTeC using the Structural Similarity Index Measure (SSIM). By systematically analyzing data loss and testing a diffusionbased approach, this study aims to enhance the reliability of singleimage texture completion techniques, contributing to improved 3D face modeling for applications in biometric security, virtual reality, and digital human representation.