FaceMoE: Mixture of Experts for Low-Resolution Face Recognition

2026-06-30

AI Summary

This work addresses the challenge of low-resolution face recognition, where identity information is often degraded due to blur, occlusion, and domain shift, causing conventional single-encoder models to suffer from poor generalization and catastrophic forgetting of pretrained knowledge. To overcome these limitations, the authors propose FaceMoE, the first approach to integrate a Mixture of Experts (MoE) architecture into low-resolution face recognition. Built upon a MoE Transformer with Top-k routing, FaceMoE dynamically activates multiple specialized feedforward experts to perform resolution-aware sparse feature extraction. Each expert autonomously specializes in distinct facial semantic regions, thereby enhancing model capacity while preserving pretrained representations. Extensive experiments demonstrate that FaceMoE significantly outperforms state-of-the-art methods across 11 diverse face datasets spanning high, mixed, and low resolutions.

Abstract

Low-resolution face recognition (LR-FR) remains a challenging task due to poor feature extraction and aggregation, as probe images often contain limited identity information resulting from extreme degradations such as blur, occlusion, and low contrast. Additionally, the domain gap between high-resolution (HR) gallery images and low-resolution (LR) probe images poses a significant challenge. A single feature encoder struggles to generalize effectively across both domains when fine-tuned on an LR dataset, and this issue is further magnified by catastrophic forgetting. To address these challenges, we propose FaceMoE, an effective adaptation of Mixture of Experts (MoE) transfomer architecture for low-resolution face-recognition . Specifically, we introduce multiple specialized feed-forward network (FFN) experts and incorporate a top-k router, which dynamically assigns tokens to appropriate experts. This design emergently promotes specialization across experts for different semantic regions of the face, which enables FaceMoE to perform resolution-aware feature extraction. Moreover, the top-k router facilitates sparse expert activation, enabling the model to preserve pretrained knowledge when finetuned on a LR dataset, while increasing model capacity without proportional computational overhead. FaceMoE is trained with a combined face recognition loss, router z-loss, and load balancing loss to ensure expert specialization and stable training. To the best of our knowledge, this is the first work leveraging MoE for LR-FR. Extensive experiments across eleven datasets, spanning HR, mixed-quality, and LR benchmarks, demonstrate that FaceMoE significantly outperforms state-of-the-art methods. Code: https://github.com/Kartik-3004/FaceMoE

Problem

• low-resolution face recognition
• feature extraction
• domain gap
• catastrophic forgetting
• identity information

Innovation

• Mixture of Experts
• Low-Resolution Face Recognition
• Sparse Activation
• Resolution-Aware Feature Extraction
• Catastrophic Forgetting Mitigation