An Energy-Guided Distribution Calibration Method for Open-Set Few-Shot Face Recognition

Open-set few-shot face recognition (OSFSFR) requires both strong generalization from scarce samples and reliable rejection of unknowns. Existing methods discard base-class knowledge after pre-training, while pseudo-labeling suffers from noisy labels and base–novel distribution gaps. We propose the BEDCM, a unified framework with four strategies: (1) consistency-based pseudo-label quality assessment that filters high-confidence samples and repurposes low-quality ones as unknown-augmented signals; (2) distribution calibration that generates intra-class virtual features from base-class statistics; (3) boundary-unknown augmentation that synthesizes challenging samples between class prototypes to form an out-of-distribution set; and (4) lightweight prior logit correction to mitigate class imbalance. An energy margin constraint widens the energy separation between known and unknown-augmented samples during training. Experiments on CASIA and IJB-C show that the BEDCM achieves competitive performance at strict FARs (0.1% and 1%), with consistent gains over existing methods. Threshold perturbation analysis further confirms reduced threshold sensitivity.