EUCB-Net: efficient underwater image enhancement with convolutional boosting

The enhancement of underwater images is vital for diverse fields such as marine biology and underwater robotics. However, it is often impeded by the intricate optical properties of water, which lead to distortion, reduced contrast, and color degradation. Although conventional techniques are computationally economical, they are not adequately effective in tackling these issues. While deep learning approaches hold promise, they depend heavily on paired training data, which can be challenging to acquire. To overcome these constraints, we introduce EUCB-Net, an unsupervised methodology that harnesses multi-scale convolutional operations and an Enhanced Up-Convolution Block (EUCB) for efficient image enhancement. EUCB-Net incorporates uncertainty modeling to enhance images adaptively, accounting for the natural ambiguity in the enhancement process. It also utilizes a Multi-scale Convolutional Attention Module (MSCAM) to further refine feature maps. Our experiments across various public datasets show that EUCB-Net achieves the best performance to date, with marked enhancements in both quality and computational efficiency. The proposed framework is highly appropriate for real-world underwater applications, providing robust enhancement without the dependence on paired training data.