An underwater object detection method based on generative degradation simulation and hard-sample learning

Underwater images are affected by spectral attenuation, particulate scattering, and nonuniform illumination and therefore often exhibit color shifts, reduced contrast, blurred edges, and local darkening. Existing public datasets provide limited coverage of degradation types and intensities, while frame extraction from continuous videos introduces many similar backgrounds, making it difficult for detection models to fully learn object variations in complex waters. To address these problems, an underwater object detection method based on generative degradation simulation and hard-sample learning, termed GDH-Det, is proposed. The method extracts degradation conditions related to water color, turbidity, and blur level from real images. A conditional diffusion network combines these conditions with bounding-box masks to generate degraded images at continuously varying intensities, and candidate samples suitable for training are selected using structural constraints on object regions and detection-based credibility scores. During detector training, sample difficulty is computed from classification uncertainty, localization deviation, and perturbation response, the sampling probability is then determined jointly by the generation credibility score and the training epoch, allowing difficult and credible samples to receive more training opportunities. Degradation conditions further participate in multi-scale feature computation, while a paired-feature loss preserves class-related information between object regions in real images and their corresponding generated images. Experiments are conducted on four public datasets, URPC2020, DUO, UODD, and Brackish. GDH-Det achieves [email protected] values of 85.3%, 79.6%, 66.2%, and 84.7%, respectively. As degradation severity increases, GDH-Det shows a smaller performance drop. At degradation level 5, [email protected] on URPC2020 remains 74.8%, 6.7% points higher than AGW-YOLOv8. In the ablation study, the complete method improves by 4.9% points over the baseline trained only on real samples. GDH-Det contains 31.8 M parameters and requires 18.7 ms per image for inference. The experimental results show that the method maintains high detection accuracy in turbid, low-light, dense small-object, and partially occluded scenes and exhibits good adaptability under complex underwater imaging conditions.