The challenge was hosted by the New Trends in Image Restoration (NTIRE) Conference in conjunction with the 2019 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), the world’s premier vision conference, according to CVPR’s website. In addition to the image dehazing challenge, in which the Penn State teams placed first and second, the NTIRE image restoration and enhancement challenges have several other tracks, including image denoising, image super-resolution, image enhancement and image colorization.
Image dehazing involves taking an image that is hazy or foggy and making it much clearer in a matter of seconds using a mathematical algorithm. The participating teams submitted their algorithms to the judges, who then applied the algorithms to a set of hazy images to test for quality and speed of the dehazing. Of the 275 teams from universities, research institutions and corporations who competed, only 22 were selected as the top teams to be invited to compete further in the challenge.
In addition to Monga, the first-place team consisted of Tiantong Guo, Xuelu Li and Venkateswararao Cherukuri, all electrical engineering doctoral candidates. The second-place team, also advised by Monga, consisted of Guo and Cherukuri.
“In a departure from most black-box learning methods, the approach developed by our group develops novel domain-enriched techniques that integrate physical modeling of the haze phenomenon into deep learning frameworks,” explained Guo.
The Penn State researchers’ contributions to the advancement of image dehazing have several important applications. Perhaps most obviously, their image dehazing algorithm will enhance the quality of photography.
“Let’s say you go to the Statue of Liberty and take a picture, but it’s a rainy day so you capture a very hazy image,” said Monga. “Now you can process it easily to get something much better.”
Beyond helping professional and amateur photographers improve the quality of their images, the dehazing algorithm can play a role in increasing security.
“What if you have surveillance cameras in a public area, and a crime happens on a hazy, foggy day? That means that now the camera and the images from the camera are useless,” said Monga. “With this [dehazing algorithm], that’s not necessarily the case. You can still get value out of it because now, any time you have a hazy image captured, you can clean it up.”
The dehazing algorithm also has strong implications for the future of autonomous driving.
“This is going to be one of the key technologies that enables future self-driving cars,” said Monga. “As cars are driving, images are being captured of different kinds, and the car is making decisions about what objects and obstacles are there, and then making adjustments accordingly. So when an autonomous vehicle drives on a rainy day, the quality of those images suffers.”
Guo, Li and Cherukuri will attend the 2019 CVPR Conference from June 16-18 in Long Beach, California, to present their work and receive their awards.
This research was funded in part by a grant from the National Science Foundation.