Pitch-Angle Face Frontalization Using ROI- Constrained GAN for Reliable Identity Verification in Surveillance Systems

Recent advancements in face identification systems have primarily focused on frontal face images, which limits their effectiveness in real-world scenarios where faces are often captured from various angles. To address this, face rotation techniques transform multi-view images into frontal views, enabling accurate identity matching with existing databases. These techniques involve three primary rotation angles—roll, yaw, and pitch—with yaw and pitch presenting significant challenges due to occlusions. While Generative Adversarial Networks (GANs) have shown success in yaw-angle rotation, pitch-angle frontalization remains less explored due to facial asymmetry and occlusions. Our research introduces a novel GAN-based framework that incorporates a Region of Interest (ROI) loss to enhance pitch-angle face frontalization. By employing a dual-loss mechanism that combines a discriminator for image realism with an identity classifier for preserving subject identity, our approach generates realistic, identity-preserving frontal images from pitch-angle inputs, as demonstrated by experimental results.