By Sandra Cabangon

Haze does not only cause health issues such as irritation to the eyes, nose, or the throat; it also reduces the quality of images we capture. Hazy images that are supposed to help us find clues in solving cases and problems, and preserving precious moments are of little help in achieving these objectives. However, thanks to a novel technology known as "dehazing," we are now able to remove haze from the photography of a scene in order to enhance or restore image quality. Dehazing methods are constantly being developed, and one of the recent innovations is that of Neil Patrick Gallego and his colleagues Joel Ilao, Macario Cordel II, and Conrado Ruiz Jr, from De La Salle University (DLSU).

Photo Courtesy of DSLU GAME Lab

Del Gallego and his colleagues’ idea is a new approach for training a physics-based dehazing network. RGB images and depth maps – image channels that contain information relating to the distance of the surfaces of scene objects from a viewpoint – are gathered from a 3D virtual environment. Since 3D images contain depth buffers – a type of data buffer used in computer graphics to represent depth information of objects in 3D space – full image depths can be extracted using a custom shader.

Complicated as these may sound, these steps simply mean extracting valuable information to visualize the appearance of objects concealed or hidden behind the fog. A style transfer strategy, followed by unlit image prior extraction from the synthetic images, allows the dehazing network to perform effectively on real-world hazy images.These unlit image priors can also be extracted from the virtual environment. The training is a supervised image-to-image translation task, using the DLSU-SYNSIDE (SYNthetic Single Image Dehazing Dataset). DLSU-SYNSIDE consists of clear images, unlit image priors, transmission, and atmospheric maps. Pre-trained models include a style transfer network, unlit network, airlight and transmission estimators. 

They claim that this approach makes their three-stage dehazing network easier to train, as compared to unsupervised approaches. Moreover, their approach has been shown to be on par with state-of-the-art dehazing works using I-Haze, O-Haze, and RESIDE; even though their network does not make use of real-world images during training.

This novel project, which is supported and funded by De La Salle University (DLSU), Department of Science and Technology (DOST), and the Google Cloud Research program, serves as a proof of Filipinos’ talents in engineering technology. As more similar innovations emerge, there is hope for the improvement and recognition of STEM related endeavors in the Philippines.