Atmospheric haze is the bane of photographers—fine particles in the air scatter and absorb light, distorting colours, reducing clarity, and hindering the capture of high-resolution images.
Model-based dehazing algorithms can counter the obscuring effects of haze for photography, surveillance, medical imaging, and other applications, but these computational tools aren’t perfect. Zhengguo Li, a Senior Principal Scientist at A*STAR’s Institute for Infocomm Research (I2R), explains the two major hurdles in the way of the current gold standards.
“First, it’s hard for algorithms to distinguish between the true colours and brightness of objects and their hazy appearances,” said Li. “Secondly, images of the sky often suffer from amplified graininess, or noise,”
Existing tools, such as the dark channel prior (DCP) and haze line prior (HLP) models either overcompensate by introducing unwanted image artifacts, or have poor spatial coherency, meaning that they misinterpret similarities between neighbouring parts of an image.
Together with Haiyan Shu, a Principal Scientist at I2R, Li collaborated with Chaobing Zheng from the Wuhan University of Science and Technology, China, to bridge these gaps and create more effective automated image dehazing workflows. Using the DCP algorithm as a foundation, they created a new prior they termed the dark direct attenuation prior, or DDAP. The team’s innovative approach was designed to significantly reduce the ambiguity between objects shrouded in haze while maintaining crisp depictions of the sky.
However, DDAP tended to introduce undesired visual objects into the processed image. To fix this, Li and colleagues leveraged a novel concept of haze line averaging, a supplementary algorithm that helps reduce the distortions caused by DDAP by applying a weighted guided image filter with a smaller radius.
Next, addressing the problem of noise (random variations in clarity and sharpness) the researchers incorporated mathematical techniques known as Laplacian and Gaussian pyramids into their workflow. These break the image into layers, each representing different scales of resolution. “This division helps to isolate and effectively eliminate noise, enhancing clarity, especially in sky areas,” commented Li.
The authors conducted a series of experiments to compare their proposed algorithm with existing dehazing methods and found that their refined approach generated sharper, higher-quality processed images.
Looking ahead, Li is leading the team to explore new horizons in image dehazing by merging the strengths of model-based algorithms with data-driven methods. “Our initial results are promising and we have already earned a patent on our algorithm,” shared Li.
The A*STAR-affiliated researchers contributing to this research are from the Institute for Infocomm Research (I2R).
