Compressed-Domain Temporal Cues and Degradation-Aware Distillation for Lightweight Dynamic Facial Expression Recognition on Edge Devices

Dynamic facial expression recognition (DFER) has advanced rapidly, but leading systems use transformer backbones with more than 80 million parameters and assume fully decoded, high-quality video. Real deployments instead deliver small, heavily compressed and partially occluded faces to hardware with a few-watt power budget. We address this gap in three ways. First, DFEW-D and FERV39k-D form a reproducible degradation protocol that jointly sweeps face resolution, codec constant rate factor and occlusion over five severity levels and three codecs. Second, CoDA-Net is a 6.8 M-parameter network that reads motion vectors and residuals from the coded bitstream, combines them with appearance features from sparse intra frames and aggregates the sequence with a bidirectional state-space core, without reconstructing inter-coded frames. Third, quality-conditioned distillation uses a small quality head to estimate per-frame degradation and scales the supervision from a frozen teacher accordingly. We evaluate on DFEW, FERV39k and MAFW. On DFEW-D, CoDA-Net retains 84.2% of its reference-encoding weighted average recall at the most severe degradation level, against 66.7% for the strongest baseline, and runs at 32.3 clips/s on a Jetson Orin Nano and 5.4 clips/s on a Raspberry Pi 5. This suffices for stride-one prediction on the Jetson and for non-overlapping windows on the Pi 5. CoDA-Net is thus a robustness–efficiency trade-off for edge hardware, not a replacement for server-scale models.