TST-YOLO: TOKEN-SELECTIVE TWIN-ENHANCED YOLO FOR ROBUST UNDERWATER FISH DETECTION AND PHENOTYPE ESTIMATION ON AUVS

Underwater fish detection remains difficult due to turbidity, color cast, motion blur, and the prevalence of small, fast, look-alike targets on embedded AUV hardware. We present TST-YOLO, a compact detector that combines four synergistic components: (i) a physics-aware digital-twin synthesizer that exposes the model to realistic water-optics shifts during training; (ii) a lightweight pre-enhancement fusion stage that mitigates color/contrast bias before inference; (iii) a token-selective transformer head that prunes and merges low-information tokens for AUV-grade efficiency; and (iv) a phenotype-guided auxiliary head that estimates lateral-line scale counts to regularize features toward biologically meaningful structure. Evaluated under identical training budgets on DeepFish and DePondFi’23, TST-YOLO improves mAP@50–95 by +6.1 and +6.0 points, respectively, over strong YOLOv8/YOLOv7 baselines, with +5.5 APS_SS gains on small targets. Confidence calibration also improves (ECE ↓34%, Brier ↓), while the token-selective head reduces transformer tokens by ≈30% at equal or better accuracy, cutting end-to-end Jetson latency by ~6%. Results are reported as 5× runs (mean±SD) with bootstrap confidence intervals, paired tests, and McNemar analyses. Beyond accuracy gains, the contribution is positioned at the systems level, demonstrating how domain-aware data synthesis, reliability-oriented evaluation, and token-efficient transformer design can be jointly integrated for trustworthy, resource-constrained intelligent perception. This emphasis on efficiency, robustness, and statistical rigor highlights the relevance of TST-YOLO as an information-technology solution for dependable autonomous sensing rather than a task-specific detector alone.