Enhanced Human Face Recall by Reconstruction of Real and Synthetic Occlusions

Abstract

This study investigates occluded face recognition using a Complete Face Recovery Generative Adversarial Network (CFR-GAN). It directly addresses the complexities of identifying faces obscured by real-world occlusions, while underlining the critical role of high-quality input data. The CFR-GAN is trained on both occluded and rotated face datasets, with synthetic occlusions applied for data augmentation. In Experiment 1, the model achieved a poor result of 15.80% accuracy for reconstructing images with sunglasses and 24.11% for masked faces, but produced a consistent recall rate of 100%. Experiment 2 utilised the Labelled Faces in the Wild (LFW) dataset, where the CFR-GAN yielded an accuracy of 52.27%. Despite improvements with higher-quality inputs, the model introduced artefacts and misaligned features, underscoring the ongoing difficulties in occluded face recognition. The results underscored the importance of data quality, advanced training techniques, and refined evaluation metrics in addressing these real-world challenges. While this work demonstrates promise, further study is required to translate these findings into practical applications in biometric security and related areas.