A Multi-Branch Feature Builder and Fusion Architecture for Robust Zero-Shot Face Recognition

This technical report presents a multi-branch feature builder and fusion architecture for robust face recognition. The proposed system extracts complementary representations through multiple feature branches, concatenates the resulting features, and applies a dedicated fusion module before global average pooling and identity embedding. The embedding is trained using an ArcFace objective. The system is evaluated in a zero-shot setting on LFW, with LFW not used for training or fine-tuning. Additional experiments examine cross-detector generalization and robustness under compound perturbations including rotation, shear, blur, color/BGR variation, and the absence of face alignment. A qualitative masked-face evaluation is also presented. The report documents the architecture, training methodology, evaluation protocol, robustness experiments, feature visualizations, ablation considerations, limitations, and reproducibility requirements. Reported results include approximately 99.95% zero-shot LFW verification accuracy, approximately 96% cross-detector performance, and approximately 85–90% performance under the evaluated compound perturbation setting. This work is presented as an independent research contribution and is intended to provide a reproducible basis for further investigation into multi-branch feature representations and robust face recognition.