FACE RECOGNITION IN MASKED SCENARIOS BY USING CYCLE GANS

Abstract

The COVID-19 pandemic’s widespread use of face masks has critically disrupted traditional Face Recognition (FR) systems, which rely on unobstructed facial features. Masks occlude key regions (eyes, nose, mouth), leading to security vulnerabilities, authentication failures, and a surge in false positives. To address this, we propose a hybrid framework combining Histogram of Oriented Gradients (HOG) landmark extraction with Cycle Consistent Generative Adversarial Networks (CycleGANs) for masked face reconstruction. By preprocessing the CelebA dataset and employing HOG-guided spatial mapping, our method in paints occluded regions while preserving facial structure. Quantitative evaluations demonstrate strong performance, achieving a Peak Signal to Noise Ratio (PSNR) of 28.5 dB and Structural Similarity Index (SSIM) of 0.89, alongside reduced cycle consistency errors (L1 loss: 0.021). The framework outperforms traditional inpainting methods in realism (82% user preference) and maintains robustness across mask types, poses, and lighting. Computational efficiency (98 MB model size, 45-minute epochs on T4 GPUs) supports real-world deployment, while ethical safeguards privacy-aware protocols and bias mitigation—address risks of misuse and demographic disparities. This work bridges the gap between FR limitations and pandemic-era realities, offering a secure, adaptive solution for occlusion-resilient biometric systems, with a 94% recognition accuracy.