Abstract
Underwater object detection (UOD) is significantly hampered by the physical properties of aquatic environments, which cause severe light absorption and scattering. Traditional approaches often rely on image enhancement as a preprocessing step or direct retraining on degraded data, both of which suffer from computational inefficiency or poor feature extraction. This paper introduces a novel Residual Feature Transference Module (RFTM) designed to bridge the statistical distribution gap between heavily degraded underwater images and high-quality, "detector-friendly" images. By learning a Heavily-Degraded Prior (HDP) through unsupervised feature alignment, the proposed method enhances the performance of standard object detectors without the overhead of full image restoration.
Introduction
Underwater environments present a unique set of challenges for computer vision. The selective absorption of light wavelengths and the presence of suspended particles lead to low contrast, color shifts, and significant blurring. These factors result in a substantial "feature distribution gap" between standard terrestrial datasets and underwater imagery. While Underwater Image Enhancement (UIE) techniques can improve visual quality for human observers, they do not necessarily preserve or enhance the semantic features required for machine perception. This research shifts the focus from visual restoration to feature-level alignment, specifically targeting the most heavily degraded regions of underwater scenes.
The Distribution Gap Problem
The authors identify a critical bottleneck in UOD: the performance of standard detectors drops precipitously when encountering heavily degraded (HD) patches, whereas lightly degraded (LD) patches remain relatively manageable. By analyzing the deep representations of these patches, researchers discovered a massive statistical discrepancy between HD features and those from "detector-friendly" (DF) images. This gap suggests that the primary obstacle to accurate detection is not the underwater environment as a whole, but the specific loss of structural information in heavily obscured regions.
Methodology
The Residual Feature Transference Module (RFTM)
The RFTM is a lightweight, plug-and-play component designed to be integrated into the backbone of existing detectors (e.g., ResNet). Its architecture is intentionally simple, consisting of a MaxPool layer followed by three $3 \times 3$ convolutional layers. This simplicity ensures that the module adds minimal computational overhead while effectively transforming feature maps.
Two-Stage Training Strategy
The implementation follows a distinct two-stage process:
- Stage 1: Unsupervised Transference Training: The RFTM is trained to map the deep representations of HD underwater patches to match the distributions of DF patches. This is achieved using a Kullback-Leibler (KL) divergence loss. Crucially, this stage is unsupervised and does not require manual semantic labels, as it focuses on the statistical properties of the degradation rather than the object content.
- Stage 2: Detector Fine-tuning: Once the RFTM has learned the Heavily-Degraded Prior (HDP), its weights are frozen. The rest of the detection network (the neck and head) is then fine-tuned on labeled underwater datasets to optimize for final classification and localization tasks.
The DFUI Dataset
To facilitate this training, the authors introduced the Detector-Friendly Underwater Images (DFUI) dataset. This dataset was curated by running a standard Cascade R-CNN detector across a vast collection of underwater images and selecting only those where the detector achieved high confidence (AP $\ge$ 60). This serves as a benchmark for the feature distributions the RFTM aims to emulate.
Experimental Results
The proposed method was evaluated on the URPC2020 and UODD benchmarks, demonstrating significant improvements over baseline models:
- Performance Gains: On the URPC2020 dataset, integrating the RFTM into a ResNet50 backbone improved Average Precision (AP) from 45.0 to 48.2. With a ResNetXT101 backbone, the AP increased from 45.2 to 50.9.
- Efficiency vs. Transformers: The RFTM-enhanced CNN models outperformed several state-of-the-art Transformer-based detectors, such as Swin Transformer, while maintaining faster inference speeds and a smaller parameter footprint.
- Robustness: Qualitative analysis indicates that the HDP-informed model remains effective in near-zero visibility conditions where both baseline detectors and enhancement-based pipelines typically fail.
Discussion and Limitations
A key insight of this work is that "visual quality" is not synonymous with "detectability." By focusing on the statistical alignment of features, the RFTM prioritizes the information most relevant to the neural network's decision-making process.
However, the method is sensitive to the transmission threshold ($T$) used to define heavily degraded patches. The optimal threshold varies between datasets (e.g., 0.7 for URPC vs. 0.6 for UODD), suggesting that future iterations could benefit from an adaptive thresholding mechanism. Additionally, while the RFTM is backbone-agnostic, its performance is still inherently limited by the quality of the initial features extracted by the underlying network.
Conclusion
By learning a Heavily-Degraded Prior, this research provides a computationally efficient path toward robust underwater object detection. The RFTM successfully bridges the gap between degraded inputs and detector-friendly representations, proving that targeted feature transference is a viable alternative to traditional image enhancement. Future work may extend this concept to other challenging domains, such as detection in heavy fog, nighttime surveillance, or aerial imaging through cloud cover.
Summary updated