MUFFNet: Lightweight dynamic underwater image enhancement network based on multi-scale frequency

Dechuan Kong ^1,2* Yandi Zhang ³ Xiaohu Zhao ^2* Yanqiang Wang ¹ Lei Cai ¹

• ¹ School of Artificial Intelligence, Henan Institute of Science and Technology, Xinxiang, China
• ² National and Local Joint Engineering Laboratory of Internet Application Technology on Mine, China University of Mining and Technolog, Xuzhou, China
• ³ School of Information Science and Engineering, Shenyang University of Technology, Shenyang, Liaoning Province, China

The advancement of Underwater Human-Robot Interaction technology has significantly driven marine exploration, conservation, and resource utilization. However, challenges persist due to the limitations of underwater robots equipped with basic cameras, which struggle to handle complex underwater environments. This leads to blurry images, severely hindering the performance of automated systems. We propose MUFFNet, an underwater image enhancement network leveraging multi-scale frequency analysis to address the challenge. The network introduces a frequency-domain-based convolutional attention mechanism to extract spatial information effectively. A Multi-Scale Enhancement Prior algorithm enhances high-frequency and low-frequency features while the Information Flow Interaction module mitigates information stratification and blockage. A Multi-Scale Joint Loss framework facilitates dynamic network optimization. Experimental results demonstrate that MUFFNet outperforms existing state-of-theart models while consuming fewer computational resources and aligning enhanced images more closely with human visual perception.