Visibility-Aware Patch Scheduling for Underwater Object Recognition

Abstract

Underwater object recognition is affected by light attenuation, turbidity, backscatter, color distortion, and uneven visibility. Efficient visual models are needed for autonomous underwater vehicles, but aggressive patch reduction may remove weak object cues in low-visibility scenes. This study investigates visibility-aware patch scheduling for underwater object recognition. We propose AquaPatch-Net, which estimates local visibility from contrast decay, blue-green color shift, edge attenuation, and suspended-particle noise. Clear background regions are processed through a compact convolutional path, while low-visibility object candidates are retained in a transformer path with enhanced contextual attention. A confidence recovery module further revisits uncertain patches when the initial prediction is unstable. The evaluation combined public underwater image datasets and field data collected during 27 autonomous underwater vehicle dives. The final dataset included 62,400 labeled images, 11 object categories, four turbidity levels, and depth records from 3 to 46 meters. Under clear-water conditions, AquaPatch-Net reduced average computation by 35.2% with no measurable loss in top-1 accuracy. Under high turbidity above 18 NTU, object recall reached 81.7%, compared with 74.9% for a fixed-compression transformer. In field deployment, the model processed 18.5 hours of continuous video and reduced onboard GPU energy consumption by 23.6%. Missed detections of small marine debris decreased from 17.8% to 11.3%, while false alarms from sand ripples and bubbles decreased by 14.6%. The average decision latency remained below 34 ms per frame, meeting real-time navigation requirements. These results show that visibility-aware patch scheduling can support efficient and reliable underwater perception in resource-limited robotic platforms.