Underwater Monocular Depth Estimation Combined with Physical Prior Information
Chaicheng Jiang, Haoran Zhen, Gang Wan, Sisi Zhu, Xinyu Li, Xianbo Xiang · IFAC-PapersOnLine · 2025
In this study, a deep learning-based underwater monocular depth estimation method is proposed to enhance the 3D perception capability of low-cost underwater robots. The method incorporates unique domain knowledge of imaging in underwater scenes to guide model learning. First, a transmission map containing depth information of the scene is obtained from the original RGB image using the underwater dark channel prior and concatenated with the original image to form an input space. Second, based on the characteristics of underwater light propagation, an underwater light attenuation prior loss is proposed to penalize erroneous prediction results in the background region and avoid artifacts on the depth map. Finally, a lightweight U-shaped network is designed, containing a MobileNetV2-based encoder, a neck module based on large kernel convolution, a decoder based on depthwise convolution, and a vision transformer-based regression module. Quantitative and qualitative performance comparisons on FLsea, a dataset of real underwater scenes, demonstrate that the proposed method outperforms existing methods while having a comparable number of parameters.