A Robust Physics-Aware and Attention-Driven Task-Oriented Deep Learning Framework for Underwater Object Detection

Object recognition in underwater environments remains problematic because of the strong reduction of the appearance, contrast and boundary information of objects due to light absorption, wavelength-dependent attenuation and backscatter. The effects are especially challenging for spotting small and low-contrast aquatic objects under murky underwater conditions. The existing methods generally either decouple image enhancement and object detection or use unrestricted enhancement networks, which may lead to objective mismatch between perceptual restoration and detection performance. To overcome these restrictions, in this work we present a task-oriented deep learning framework, named PA-AFNet (Physics-Aware, Attention-Fused Network), which combines the underwater image generation model with the multi-scale object detection. The proposed system consists of a differentiable physics-aware enhancement module, which estimates the transmission and backscatter parameters and reconstructs a physically restricted improved image. The estimated transmission information is kept as an auxiliary signal and directly fed into a multi-scale attention fusion module that is integrated with the YOLOv11 feature pyramid. This allows the detector to adaptively refine the features according to the spatially variable degradation characteristics of the underwater images. The whole framework is optimized in an end-to-end manner under the detection objective, such that the enhancement and detection modules can be learned collaboratively rather than separately. The experiments were conducted on URPC2017 and URPC2018 underwater object detection benchmarks with the same training setting. The proposed framework is effective with 88.2% precision, 84.1% recall, 88.2% mAP@0.5, 54.8% mAP@0.5:0.95 and an F1-score of 86.10% to demonstrate the combination of physical image modeling, task-oriented enhancement and degradation-conditioned multi-scale feature attention for underwater object detection.