Biomimetic polarization vision target detection method for underwater autonomous vehicles

Underwater visual target detection is critical for AUVs but constrained by light attenuation, scattering, and low visibility, where traditional algorithms lack accuracy and robustness. Inspired by mantis shrimp polarization perception, this paper proposes a biomimetic method integrating three core components: Sea-thru for polarization image restoration (mitigating backscattering/attenuation, PSNR=26.77 dB, UIQM=3.59), a PGA module for fusing four polarization angles (0°,45°,90°,135°) with reduced complexity, and the ShareCMP-TOOD framework (multi-modal weight sharing + taskaligned detection). Experiments on the UP- Light dataset show ShareCMP-TOOD achieves mAP@0.50:0.95=0.747 (+3.2% vs. original TOOD), enhancing AUVs’ target detection in complex underwater environments for marine resource exploration and environmental monitoring.