Attention Distillation with Sample Re-weighting and Adaptive Margin for Low-Resolution Face Recognition

Low-resolution face recognition poses challenges due to the loss of fine details and increased image noise. To address this, we propose an attention distillation framework enhanced with two novel modules. Firstly, the Sample Re-weighting (SRW) module assesses low-resolution sample quality by assigning importance weights based on image quality. This reduces the impact of low-quality samples during model training, enabling better handling of noisy samples. Secondly, the Adaptive Margin Module (AM) addresses intra-class variation by introducing category-dependent learnable parameters to adjust boundaries for each class. This captures variations within categories, resulting in improved recognition performance, especially for classes with limited samples. We evaluate our framework on AgeDB-30 for face verification and Tinyface for face recognition, demonstrating significant improvements at lower resolutions. Our method surpasses the state-of-the-art A-SKD approach, particularly in handling challenges posed by low-resolution images.