Exploring the Depth From EAST: Efficient Aggregated State-Space Tanh-Tuned Model for Underwater Object Detection

Abstract

Underwater object detection faces severe challenges due to light attenuation, color distortion, and low contrast. This letter presents EAST-YOLO, an efficient architecture achieving 79.6% average mAP@0.5 across six underwater datasets—2.1% higher than YOLO11n—while maintaining 2.6 M parameters, 6.5 GFLOPs, and 70 FPS real-time performance. Three problem-driven modules address specific underwater challenges: VSS-Enhanced Block for visibility-limited global context modeling with $\mathcal {O}(N)$ complexity, Aggregated Pathway Block for refraction-robust multi-scale detection, and Tanh-Tuned Attention Block for spatially-adaptive feature modulation. Extensive evaluation on RUOD, DUO, URPC2020, UTDAC2020, DFUI, and AUDD datasets demonstrates EAST-YOLO’s effectiveness as a practical solution for resource-constrained underwater applications, with promising robustness across diverse degraded conditions.