Underwater images suffer due to light scattering and color absorption, which is challenging for marine researchers, underwater heritage miners, and aquatic robotics. Due to the unavailability of ample real-world underwater ground truth images, it becomes more challenging to frame a deep learning-based underwater image enhancement model to give clear underwater pictures in the existing models. Existing datasets use synthetic images to generate ground truth. In this paper, we propose a new dataset with 3000 real-world reference images taken at depths up to 5 meters so that the collected images are clear and serve the purpose. For raw images, we propose a new algorithm that considers the parameters for color channels for varying noise and color absorption levels, from clear to turbid. We further propose a CNN-based model named UMLDnet with minimum loss dehazing using scene radiance from background light and transmission map. Compared to the state-of-the-art methods, UMLDnet can generate enhanced images with better PSNR and a much lighter deep-learning network.

Fig. 1: Sample images of dataset AUIED3K. first row: Reference, second row: Raw images

Fig. 2: An overview of UMLDNet architecture

Praveen Saini, Navjot Singh, Anshu S. Anand, "AUIED3K: A New Andaman Underwater Image Enhancement Dataset for Deep Learning-Driven Image Enhancement with Minimum Loss Dehazing," 50th IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2025. (Accepted)

Google Drive Link