Enhancement as Augmentation: Improving Detection in Highly Degraded Underwater Images Through Mixed-Domain Training

Ashraf Saleem, Alex Hromada, Ali Awad, Amna Mazen, Evan Lucas, Timothy C. Havens; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 1096-1105

Underwater object detection is challenged by visibility degradation caused by absorption, scattering, and turbidity. Although underwater image enhancement (UIE) is often used as a preprocessing step, prior work shows that enhancement can distort appearance and reduce detector performance. We revisit UIE from a data augmentation perspective and propose a mixed-domain training framework in which original images are paired with enhanced variants generated by four state-of-the-art UIE models: ACDC, AutoEnhancer, TUDA, and Semi-UIR. This design isolates the effect of enhancement-induced domain shifts while keeping the detector architecture and inference pipeline unchanged. Experiments on a curated subset of the USGS round goby dataset reveal that perceptual enhancement quality does not predict detection effectiveness: the enhancer with the lowest UIQM, UCIQE, and CCF scores yields the strongest mAP improvement, whereas the highest-scoring enhancers produce the weakest detection results due to overenhancement. We benchmarked the evaluation metrics across the original, enhanced, and mixed-domain datasets, and observed that Mixed-domain training compensates for the effects of enhancement and consistently improves mAP@50, F1-score, and detection accuracy, demonstrating that enhancement-as-augmentation is an effective and lightweight strategy for improving the robustness of underwater object detection.