Hierarchical Neuron Energy Feature Decoupled Distillation for Low-Resolution Face Recognition

Low-resolution face recognition poses significant challenges in real-world applications, particularly for deployment on resource-constrained mobile terminals where both accuracy and efficiency are critical. While knowledge distillation has emerged as a promising solution, the existing methods suffer from performance degradation due to negative knowledge transfer when significant capacity gaps exist between teacher and student networks. To address the issue, we propose a novel Hierarchical Neuron Energy Feature Decoupled Distillation (HNEFDD) framework. Firstly, to mitigate the semantic feature gap between teacher and student networks, we integrate a neuron energy feature fusion transfer module into the student network. This module effectively aligns the features of teacher and student networks, alleviating negative knowledge transfer. At the same time, it can adaptively assign higher weights to discriminative facial regions, enabling the network to enhance locally robust structural information and extract semantically richer features in low-resolution face images. Secondly, to achieve efficient and accurate knowledge transfer, we employ a logit-standardized decoupled distillation mechanism that separates the distillation loss from the intermediate-layer auxiliary branches of the teacher and student networks into target-class and non-target-class components. This strategy enables the student network to more accurately learn both the key semantic representations of target-class features and the discriminative boundaries among non-target-class features, thereby achieving efficient and accurate knowledge transfer. Extensive experiments conducted on low-resolution benchmark datasets demonstrate the superiority of the proposed HNEFDD method, which outperforms the current state of the art (SOTA).