Joint Deep Learning and Atmospheric Light Scattering for Fast Image Dehazing

Abstract

Haze significantly degrades image and video quality by reducing contrast and visibility. A widely adopted solution is the Atmospheric Light Scattering (ALS) model, which requires estimating two key unknowns: atmospheric light and the transmission map. To estimate the depth map, we employ a lightweight encoder-decoder deep model and, rather than predicting a single air light value for each image channel, we generate an air light map. By this way, for each pixel of the hazy image we can model the real-world hazing effect more properly. This approach leverages the power of both Deep Learning (DL) and traditional methods. Our method decreases the processing load of depth estimation by using a lightweight model. In addition, it expands the air light estimation to all image pixels. This allows our method to differentiate itself from existing works. By keeping the visual quality of the dehazed imagery, our method, on a single GPU, achieves 10 fps and 17 fps for 480p and 360p image resolutions, respectively. Benchmark comparisons demonstrate that this method is competitive with existing state-of-the-art real-time dehazing methods in terms of efficiency and visual quality.