Foreground Probability Enhancement for Accurate Underwater Object Detection

Underwater object detection is essential for marine environmental protection and aquaculture. However, challenges such as low contrast and insufficient illumination in underwater environments can significantly impact detection accuracy. Although underwater image enhancement methods have been proposed to improve image quality, relying solely on enhanced images may inadvertently distort or remove key features, which can degrade model performance. Multi-task learning approaches attempt to integrate tasks like denoising before the detection task, but these methods introduce additional training data and time. To address these challenges, we propose the Foreground Probability Enhancement Network. The core component, the Foreground Probability Enhancement module, enhances foreground features by estimating foreground probabilities, refining the feature map, and suppressing background noise through weighted fusion. The refined feature maps are fed into a transformer encoderdecoder for robust object detection. The Foreground Probability Enhancement Network achieves state-of-the-art performance on the DUO and RUOD datasets, improving detection accuracy and efficiency without relying on additional preprocessing or external datasets.