Advancing monocular depth estimation by integrating underwater optical imaging priors into transformer-based network

Underwater monocular depth estimation is highly valuable for resource-constrained miniature autonomous robots due to its compact installation and low cost. However, underwater images are often degraded by light scattering and absorption, resulting in poor quality, low contrast, and loss of detail, making underwater monocular depth estimation even more challenging. A key research challenge is leveraging the unique characteristics of underwater imaging to improve depth estimation accuracy. To address this issue, we propose an underwater light priors processing module (ULPM), which integrates with features extracted by EfficientNet and is subsequently fed into a transformer-based network for depth estimation. It incorporates a physical model of underwater optical imaging into the transformer network, enhancing the accuracy of depth estimation. Additionally, we introduce a classification loss during training, leveraging the dual advantages of treating monocular depth estimation as both a regression and a classification task, thus enabling the network to generate more robust and accurate results. Experimental results demonstrate that our algorithm achieves the best performance on the Flsea and USOD10K datasets, outperforming existing methods across multiple evaluation metrics.