Turbid water no longer hides the catch: diffusion-generated degraded samples keep detection at 74.8% under heavy degradation
Synopsis
Underwater images suffer color shifts, low contrast and blurred edges from light attenuation, particle scattering and uneven illumination, while public datasets cover only a limited range of degradation types and intensities. GDH-Det extracts water color, turbidity and blur level from real images as degradation conditions, uses a conditional diffusion network with bounding-box masks to generate degraded images at continuously varying intensities, and screens candidates by object-region structure and detection credibility. During detector training, sample difficulty is computed from classification uncertainty, localization deviation and perturbation response to set sampling probabilities, reaching mAP@0.5 of 85.3%, 79.6%, 66.2% and 84.7% on URPC2020, DUO, UODD and Brackish.
Interpretation
Training data no longer comes only from limited waters and consecutive video frames; degraded images are generated at continuously varying intensities from water color, turbidity and blur level while keeping the original class and bounding-box annotations. Public underwater datasets are usually collected from limited waters or continuous video, yielding many similar backgrounds and a narrow range of water color, turbidity and blur variation, and fixed-parameter processing cannot extend degradation types or intensities. On URPC2020, training on real samples alone gives 80.4% mAP@0.5, adding conditional-diffusion degraded samples raises it to 82.5%, and adding credibility filtering reaches 83.3%.
Generated samples are not folded into the training set once; their use frequency changes with training progress and sample difficulty, and low-credibility generated images are restricted. With random sampling in every epoch, samples already recognized correctly repeatedly occupy batches, while samples with ambiguous classification, large localization error or sensitivity to perturbation get too few training opportunities. On URPC2020, adding stage-wise sample scheduling raises mAP@0.5 from 83.3% to 84.3%, above Focal loss at 83.7%, OHEM at 83.9% and difficulty-based loss reweighting at 84.0%.
Degradation conditions enter multi-scale feature computation, and features of the same object region in a real image and its generated counterpart are pulled together, so image-quality information participates in detection feature extraction. Earlier detectors mostly handled low-quality images by modifying input quality or detection architecture, without letting degradation information directly weight multi-scale features. In the URPC2020 ablation, degradation-aware features lift mAP@0.5 from 84.3% to 84.9%, and adding the paired-feature loss reaches 85.3%, a cumulative 4.9 points over the real-sample baseline.
Perspective
The method targets underwater object detection and is validated on four public datasets, URPC2020, DUO, UODD and Brackish, with training, validation and test splits separated by data source so consecutive frames from one video do not cross subsets. Only the detector is kept at inference, the diffusion generator is not loaded, and inference takes 18.7 ms per image with 31.8M parameters. In cross-dataset experiments the model is trained only on URPC2020 and evaluated directly on other datasets without parameter updates.
On 96 frames with strong localized illumination or saturated light beams, GDH-Det reaches 61.7% mAP@0.5, below AGW-YOLOv8 at 62.3%; for objects smaller than 16×16 pixels its recall is 69.8%, also slightly below 70.4%. Degradation conditions come from three image-level statistics, giving limited sensitivity to strong spatial illumination variation, and extremely small objects carry little regional texture, weakening the paired-feature constraint. When the generated-to-real sample ratio reaches 1.5, mAP@0.5 falls to 84.5%, and the optimal credibility threshold across the four validation sets ranges from 0.60 to 0.64.