Skip to main content

Enhancing underwater images thanks to AI

May 29, 2024

Side-by-side comparison of two fish images: the left image shows a fish with dull, blue-green hues, while the right image shows the same fish with restored vibrant yellow and black colors, demonstrating the AI's ability to enhance underwater image clarity and color. New AI restores vibrant colors to underwater images, transforming the original on the left into the image on the right. (Source: IEEE Transactions on Broadcasting)
M. Omair Ahmad

Imagine diving into the ocean and capturing stunningly clear photographs of marine life, or broadcasting underwater explorations with vivid, high-quality visuals.

A new study published in IEEE Transactions on Broadcasting introduces a novel approach to enhance underwater images.

This work was carried out by PhD student Alireza Esmaeilzehi (now a postdoctoral fellow at the University of Toronto) and visiting researcher Yang Ou (now at Chengdu University), under the supervision of Professors M. Omair Ahmad and M.N.S. Swamy, both from the Department of Electrical and Computer Engineering at the Gina Cody School of Engineering and Computer Science.

The team’s method leverages deep learning, a type of artificial intelligence that uses neural networks to mimic the human brain's ability to learn from data.

The method integrates information from various sources, including underwater medium transmission maps—which help understand how light travels through water—and atmospheric light estimation techniques—which assess lighting conditions. This combination creates a rich set of features that enhances the AI's ability to improve image quality.

M.N.S. Swamy

The researchers employed a three-stage training process for their deep learning model:

  1. Supervised Learning: The AI is trained using labeled data, where each image is paired with a high-quality reference, teaching the model to recognize and improve poor-quality images.
  2. Adversarial Learning: In this stage, the AI is further refined using a technique called generative adversarial networks. Here, the model learns to distinguish between real and fake images, enhancing its ability to generate realistic textures and structures.
  3. Combined Training: Finally, the model combines the knowledge gained from the first two stages, optimizing its performance to produce the best possible enhanced images.

Testing their method on various benchmark datasets, the researchers achieved outstanding results. The new technique not only improved image quality significantly but also did so with a relatively small number of parameters (2.98 million), making it efficient and less resource-intensive compared to other state-of-the-art methods that often require tens or hundreds of millions of parameters.

This research holds potential for practical applications in fields such as marine biology, where clear visuals are essential for studying marine life, and broadcasting, where high-quality underwater footage can enhance documentaries.

Learn more about the Department of Electrical and Computer Engineering at the Gina Cody School.

Back to top

© Concordia University