DNnet: A lightweight network for real-time 4K underwater image enhancement using dynamic range and average normalization
Tianyu Cao, Zhibin Yu, Bing Zheng · Expert Systems with Applications · 2025
In recent years, neural network-based methods have emerged in underwater image enhancement (UIE). Although researchers have made considerable progress in this field, applying state-of-the-art algorithms requires significant computing resources for real-time applications, making it impossible to deploy on underwater mobile devices. To address this issue, we propose the pixelunshuffle-bottleneck-pixelshuffle structure, which utilizes more parameters to improve the parallel computational efficiency and achieve higher computing efficiency. In addition, we propose fast average normalization (FAN) with a channel dynamic range (CDR) module to ensure the enhancement quality during high-speed inference. FAN uses average pooling to reduce overenhancement or underenhancement risks on channel-wise normalization. CDR can utilize dynamic range to further improve the enhanced image quality given by FAN. Moreover, we introduce a color column loss to evaluate the local color distribution between the generated image and ground truth to avoid the undersaturation problem. Our proposed method 1 outperforms existing state-of-the-art methods, with a peak signal-to-noise ratio of 26.335 and a structural similarity index measure of 0.91 on the UIEB dataset. Furthermore, our method can achieve real-time enhancement of 4K images on edge devices up to 78FPS.