UWM-Net: A Mixture Density Network Approach with Minimal Dataset Requirements for Underwater Image Enhancement

The learning-based underwater image enhancement, which is suitable for batch processing, is a pivotal research direction in underwater image processing. Extensive paired image data are required in existing learning-based methods, which necessitate considerable preprocessing and hinder the application of these methods. To address these limitations, we propose a semi-supervised approach called UWM-Net: firstly, we use a compact dataset of underwater image pairs to train the Mixture Density Network (MDN) with an underwater scene setting; subsequently, U-Net can learn underwater image enhancement more efficiently. The MDN can transform standard images into underwater scenes, reducing the reliance on paired data and making much smaller training datasets. In experimental studies, UWM-Net using only 18 pairs of underwater image data achieves highly competitive results in terms of 3 metrics compared with advanced models.