Hybrid object detection in low visibility underwater imagery

Abstract

This study presents a multilevel analysis of object detection in underwater and aerial imagery captured under limited-visibility conditions. The physical causes of input-signal degradation, including wavelength-dependent absorption, scattering, non-uniform illumination, and low-contrast blur, are systematized, feature representations that can preserve object information under low contrast are considered, and a conceptual hybrid approach to robust detection is formalized. Analysis of recent research indicates that YOLO-, Faster R-CNN-, and transformer-based detectors may lose accuracy under adverse observation conditions when low-level visual features are strongly degraded, which reduces detection confidence and increases the number of false positives. To address this problem, a conceptual multi-branch architecture is proposed that coordinates detector-oriented image enhancement, physics-guided degradation estimation, attention/context mechanisms, multiscale feature fusion, and a real-time detector based on YOLO or RT-DETR. The enhancement branch is intended to be evaluated primarily by its influence on downstream localization rather than by visual-quality metrics alone. The experimental part of this study is deliberately limited to establishing a reference baseline: YOLO11n was trained and evaluated on the single-class Common Objects Underwater (COU) dataset for underwater technical-object detection. On the held-out test split, the baseline achieved Precision = 0.889, Recall = 0.832, mAP@0.50 = 0.888, and mAP@0.50:0.95 = 0.751. These values characterize only the YOLO11n baseline and do not constitute experimental validation of the complete hybrid architecture. Accordingly, the expected contributions of the enhancement, physics-guided, attention, and fusion branches are formulated as testable hypotheses to be verified by controlled ablation studies under identical data splits and training protocols. The principal contributions of the present work are the formalization of the conceptual architecture, the definition of the baseline experimental protocol, and the specification of a validation strategy for subsequent experiments.