Novel Network Proposed to Enhance Underwater Image Quality

Jun 17, 2024 | By WANG Liusan; ZHAO Weiwei

Recently, the team led by Prof. WANG Rujing and WANG Liusan from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences, established a Learnable Full-frequency Transformer Dual Generative Adversarial Network (LFT-DGAN) to address the issue of underwater image quality degradation caused by various interferences.

The research results were published in Frontiers in Marine Science.

Underwater image enhancement technology aims to optimize the quality of underwater images and meet the diverse needs of marine scientific research, underwater robots and object recognition. Due to the unique underwater environment, noise and colour deviation often affect images, making enhancement extremely difficult. Researchers must continue exploring and innovating to improve the quality of underwater images.

This study used the knowledge of reversible convolution and adversarial neural networks to establish a dual-generative adversarial neural network model of the full-frequency transformer and verified its effectiveness by comparing multiple underwater image experimental data.

With the help of this model,researchers used image decomposition technology with reversible convolution for the first time to accurately separate the different frequency features of the image.

In addition, the study used an advanced transformer model that can learn to improve the interaction and integration of different types of information. They also created a dual-domain discriminator to better capture and analyze the frequency characteristics of the images.

"Our research results and methods have provided a solid theoretical foundation and strong support for the subsequent research and development of underwater image enhancement." said Wang Liusan.

The overall architecture of the LFT-DGAN network (Image by WANG Liusan)


Attachments Download:

Related Articles
Copyright © Hefei Institutes of Physical Science, CAS All Rights Reserved