STAR-FER: Adaptive Selective Restoration for Edge-Based Low-Resolution Facial Expression Recognition

概要

Variable interaction distances continuously degrade facial resolution in edge-based affective interaction, making low-resolution facial expression recognition (FER) difficult under limited computation. Full-face image or feature restoration incurs substantial overhead, whereas lightweight recognition pipelines may miss local cues that remain discriminative after degradation. We propose STAR-FER, a scale-and-task-adaptive selective restoration framework that balances recognition performance and resource constraints. A lightweight image-scale conditioned Gate jointly predicts a pixel-wise spatial ranking and a continuous restoration budget, then applies exact-support projection to obtain an input-dependent free-form mask. The mask-conditioned restorer retains complete facial context while concentrating sparse restoration and residual insertion on the selected support. Across seven degradation scales from x2 to x8 on the complete RAF-DB test set, STAR-FER achieves 80.92% mean accuracy and 71.64% Macro-F1, improving the low-resolution ResNet-18 baseline by 2.25 and 2.71 percentage points, respectively. It surpasses SR-FER by 1.04 accuracy points while requiring 8.84 GFLOPs, a 54.29% reduction, with 28.01% mean mask coverage, and maintains effective recognition at unseen non-integer scales. These results show that spatial-ranking and continuous-budget driven selective restoration reduces spatial modification and computation while maintaining low-resolution FER performance, offering an efficient path to resource-constrained affective interaction.

参考リンク

• Paper ↗

リンク先のライセンス・再現性・データ条件をご確認ください。自動分類や実装の目安は保証ではありません。

実装を始めるプロンプト

参考リンクを確認したうえで、次の内容をコピーして利用できます。

# 論文・記事
STAR-FER: Adaptive Selective Restoration for Edge-Based Low-Resolution Facial Expression Recognition

# 概要
Variable interaction distances continuously degrade facial resolution in edge-based affective interaction, making low-resolution facial expression recognition (FER) difficult under limited…

# 原典
https://arxiv.org/abs/2610.41310

上記の手法を実装してください。実行前に依存関係とライセンスを確認してください。