Visual quality optimization for underwater images using object perception methods

Underwater object detection remains difficult because targets are observed through wavelength-dependent attenuation, scattering, unstable illumination, and cluttered seabed backgrounds. Many existing studies improve detection either by enhancing the input image before detection or by adding generic feature-enhancement blocks to a detector. These strategies are useful, but they do not explicitly answer a task-level question: which visual changes are actually beneficial for recognizing underwater objects? A visually clearer image may still weaken object contours, enlarge background textures, or introduce enhancement artifacts. To address this problem, this paper presents PDRNet, a physics-aware and detection-oriented representation learning framework for underwater object detection. The main idea is to use enhanced images as selective training references rather than as direct detector inputs. First, a Physics-Aware Enhancement Selection (PES) module constructs a reference pool from multiple underwater enhancement operators and evaluates candidates using transmission-related reliability, background-interference risk, and detector feedback. Therefore, the model can distinguish helpful references from visually attractive but detection-unfriendly ones. Second, a Degradation-Guided Contrastive Learning (DCL) strategy transfers the selected reference information to raw-image features by pulling degraded object features toward reliable references and separating them from artifact-dominated responses. Third, Dual-Region Reference Modeling (DRR) introduces object-object structural compensation and object-background contextual suppression, allowing severely degraded objects to borrow shape cues from reliable same-category regions while reducing false alarms from seabed-like textures. Finally, a Degradation-Aware Multi-Scale Fusion (DAMF) module dynamically combines fine details and semantic features according to local degradation levels. Experiments on RUOD and URPC2020, together with a quantitative cross-dataset evaluation on UODD under a no-retuning zero-shot protocol, show that PDRNet improves detection accuracy, small-object robustness, and transfer performance without requiring multiple enhancement networks during inference. The complete RUOD model reaches 81.0% mAP, and on UODD it improves zero-shot mAP from 38.7% to 42.9%. The results indicate that modeling enhancement as a physics-aware reference signal is a more suitable direction for underwater detection than using enhancement as a purely visual preprocessing step.