Multi-scale feature modulation network introduced for advancing underwater image enhancement

Published in Journal of King Saud University - Computer and Information Sciences, their research addressed the challenge of improving image quality in underwater environments while considering the limitations of low-memory and computational power equipment.

High-quality images are important for many underwater applications, including fisheries monitoring, environmental and species conservation. However, most deep learning-based underwater image enhancement networks are not suitable for underwater equipment platforms with limited memory and computational power. This conflict poses a significant challenge in improving the quality of underwater images.

The approach in this research, which was called a simple yet effective multi-scale feature modulation network (MFMN), achieved a better trade-off between model efficiency and reconstruction performance.

"The key parts are multiscale modulation module and channel mixing module," explained Dr. WANG Liusan, member of the team.

By incorporating the multiscale modulation module similar to a visual transformer, the network extracts features from the input image and dynamically selects representative features in the image space.

To address the lack of channel feature information, a channel mixing module is introduced to enhance the spatial perspective.

Experimental results demonstrate that the MFMN method significantly reduces network parameters compared to existing techniques, making it 8.5 times smaller. Despite its smaller size, the method achieves similar performance with lower computational costs.

These findings have promising implications for applications such as underwater fisheries monitoring and environmental conservation, according to the team.

---

---