Estimation of Underwater Monocular Depth Using Laplacian Pyramid-Based Depth Residuals

With the success of deep learning, an increasing number of architectures and models have been proposed. Among these, the encoder-decoder architecture is one of the most renowned, and monocular depth estimation has also received enthusiastic research. In the current landscape of deep neural network models, two mainstream computational approaches prevail: convolution and transformer. However, the latter's substantial computational and parameter demands make it challenging to run on end devices. As a result, convolution-based research continues to evolve. To enhance the utilization of finely encoded lower-level features, we have implemented a straightforward yet highly effective approach by incorporating the Laplacian pyramid into the structure of the decoder. This entails feeding encoded features into different depth levels for deep residual decoding and progressively merging their respective outputs to reconstruct the final depth map from coarse to fine. This methodology greatly enhances accurate depth boundary estimation and global context comprehension. Furthermore, we introduce an improved activation function that enhances gradient flow, leading to faster convergence.Experimental results on benchmark datasets from various underwater environments demonstrate the effectiveness of this approach in monocular underwater depth estimation, especially when compared to existing models.