In context
Underwater robots rely on vision for tasks such as marine biology inspection and pollution monitoring, but light scattering and absorption cause severe color distortion and detail loss. By early 2025, most enhancement methods improved visual quality yet often degraded downstream object detection performance, highlighting the need for task-aware enhancement algorithms.
What was reported
Researchers from Guangxi University, Shanghai Jiao Tong University, and Sun Yat-Sen University proposed a contrastive learning-based multi-task framework that jointly performs underwater image enhancement and object detection. The method uses a region proposal module to construct positive and negative image blocks, pulling target regions closer to the original image in feature space while leveraging detection gradients to guide enhancement toward detection-friendly outputs.
Built on a cycle-generative adversarial network (CycleGAN), the approach learns clear-image features without requiring paired underwater images, reducing data demands. The generator incorporates contrastive mechanisms in residual blocks and adds a detection auxiliary task at the final convolutional layer to align features between enhancement and detection.
Experiments on EUVP, U45, and UIEB datasets for enhancement and RUOD, URPC2020, and RUIE for detection showed effective color correction while preserving structural texture. Objective metrics reached a peak signal-to-noise ratio of 24.57 dB and structural similarity of 0.88. After enhancement, detection accuracy improved by an average of 2% on Faster R-CNN and YOLOv7.
Why it mattered
This work addressed a critical gap in underwater robotics: making image enhancement serve the needs of automated perception rather than human viewing. By coupling enhancement with detection feedback, it offered a practical path to improve the reliability of vision-based underwater robots in real-world inspection and monitoring tasks.
“The detection accuracy is improved by an average of 2% on Faster R-CNN and YOLOv7 algorithms after image enhancement.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2025-01-15 · “基于对比学习的水下图像增强和检测方法”