Meningkatkan Kualitas Pengenalan Wajah Adaptif pada Wajah yang Terhalang melalui Mekanisme Perhatian Spasial

State-of-the-art face recognition models, such as AdaFace, have achieved remarkable accuracy on standard benchmarks by utilizing quality-adaptive loss functions to mitigate the impact of low-quality images. However, a significant performance gap remains in unconstrained, real-world scenarios characterized by "long-tail" spatial degradations, specifically partial occlusion and extreme pose variation. While current quality-adaptive methods can assess global image quality, they lack a mechanism to spatially discriminate between salient facial features and corrupted regions, leading to "feature pollution." This thesis investigates the hypothesis that integrating a Convolutional Block Attention Module (CBAM) into the AdaFace architecture can enhance robustness against these specific spatial challenges. We propose a novel architecture, AdaFace-CBAM, which integrates a spatial-channel attention mechanism into the ResNet-50 (ir-50) backbone. To validate the architectural design, we first conducted a rigorous ablation study which demonstrated that a targeted, late-stage integration of the attention module prevents overfitting and yields superior generalization compared to a complex multi-stage approach. The proposed model was trained from scratch on the MS1MV2 dataset and evaluated against the official AdaFace baseline across a comprehensive suite of 11 benchmarks. The experiments yielded a nuanced and significant "hybrid" result. On general-purpose benchmarks such as IJB-C and LFW, the proposed model performed at a level statistically equivalent to the baseline, confirming that the modification does not degrade core recognition capabilities. However, on specialized "challenge" benchmarks, the model demonstrated a consistent and dramatic performance improvement. Most notably, the AdaFace-CBAM model achieved a +4.82% improvement on the CPLFW pose dataset and a +1.23% improvement on the CFP-FP extreme pose dataset at a strict False Accept Rate (FAR) of 0.01%. These findings empirically validate that the spatial attention mechanism successfully filters out self-occluded regions, providing a targeted and valuable enhancement for high-security recognition in difficult, unconstrained environments.