In brief

Researchers unified the capabilities of traditional model-based and data-driven dehazing algorithms in a high-efficiency neural augmentation framework.

© Pexels

Algorithms focus on hazy images

22 May 2023

A new image dehazing algorithm sharpens both real-world and computer-generated images with unprecedented speed and accuracy.

Those who wear glasses will understand the struggle of walking outside on a cold and rainy day, trying to make sense of the world through a murky, hazed lens of fog and condensation.

Outdoor surveillance cameras, satellite imaging systems and self-driving cars all face a similar challenge: weather conditions can severely compromise image clarity. This is where dehazing algorithms can help—these computational platforms work by estimating the amount of light scattered by the haze to enhance blurry images.

Among the dehazing algorithms typically used, model-based algorithms are better suited for images captured by cameras, whereas data-driven algorithms work better for synthetic, computer-generated images. Zhengguo Li, a Senior Scientist at A*STAR’s Institute for Infocomm Research (I2R) sought to develop an integrated algorithm. “Combining the strengths of model-based and data-driven techniques could significantly reduce the number of training samples needed and improve convergence speed,” Li explained.

Together with researchers from the Wuhan University of Science and Technology, China, Li and team set about developing a high-efficiency, multi-purpose dehazing algorithm capable of processing both real-world and synthetic images.

The new algorithm uses a three-step process to restore the quality of hazy images. Processing starts by estimating the haziness and brightness of input images. This data is then funnelled into a deep-learning model that generates high quality, haze-free pictures using training data. Finally, a physical model called Koschmieder’s Law subtracts the effects of the haze by estimating how much light was absorbed by atmospheric particles.

“The resultant neural augmentation framework can improve generalisation and processing time,” said Li, adding that their innovation prevents noise amplification in images of skies, a common artifact of many dehazing algorithms.

With both accuracy and efficiency, the team’s next-generation dehazing algorithm could soon give digital imaging systems a leg up for producing consistently crystal-clear visuals, even in changing environments.

When combined with cloud computing capabilities, the dehazing algorithm could also provide photographers—amateurs and professionals alike—the ability to dehaze images directly from their smartphones. For future research, Li and colleagues are looking forward to exploring such applications further in collaboration with other teams at A*STAR.

The A*STAR-affiliated researchers contributing to this research are from the Institute for Infocomm Research (I2R).

Want to stay up to date with breakthroughs from A*STAR? Follow us on Twitter and LinkedIn!


Li, Z., Zheng, C., Shu, H. and Wu, S. Dual-Scale Single Image Dehazing via Neural Augmentation. IEEE Transactions on Image Processing 31, 6213-6223 (2022)│article

About the Researcher

Zhengguo Li was awarded a Bachelor of Science in Applied Mathematics and a Master of Engineering in Automatic Control from North-eastern University, Shenyang, China. Li completed his doctoral degree at Nanyang Technological University, Singapore. His research interests include video processing and delivery, computational photography, switched and impulsive control, sensor fusion and physics-driven deep learning. He has an extensive publication track record that spans over 200 journal and conference papers and 11 granted patents. Currently, Li is a Senior Principal Scientist at A*STAR’s Institute for Infocomm Research (I2R). He is an IEEE Fellow.

This article was made for A*STAR Research by Wildtype Media Group