Feature extraction and lightweight methods for cattle facial recognition

Current cattle-face recognition pipelines in smart agriculture still suffer from accuracy gaps and model bloat. We propose an end-to-end framework that couples enhanced feature extraction with structured compression. The backbone, DSFFNet, re-engineers ShuffleNetv2 with a hybrid parallel-attention module (MEM) that fuses spatial and coordinate cues, strengthening robustness to occlusion and pose variation. A lightweight student network, L-DSFFNet, is first slimmed via structured pruning, then refined through multi-teacher knowledge distillation. Although multi-teacher distillation has been exploited elsewhere, we quantify teacher complementarity and introduce a mask-boosted feature distillation objective tailored for cattle faces. Without architectural novelty, the pipeline delivers competitive accuracy-efficiency trade-offs on three public datasets, underscoring its practical value.