HazeAway: A Convolutional Approach to Single-Image Haze Removal

The existence of particles in the air cause haze, and this haze or fog cause degraded visibility in the captured shot from the camera. The non-uniform distribution of these particles, along with smoke, low light and pollution in the atmosphere, makes haze removal difficult in the real world images. The core computer vision tasks struggle with hazy images due to the lack of detail and poor visibility. The existing method relies on a transmission map in amalgamation with the atmospheric scattering input images to reconstruct a haze-free depiction. We suggested a single-image convolutions neural network that removes the haze present in the image and improves the perceptual quality by enhancing visibility. We used U-Net-like architecture with an encoder, bottleneck and decoder coupled with skip connections. In our experiment, we demonstrated the results on various benchmark dataset and compared our results with existing approaches. Additionally we compared the results from our network training on different image representations RGB verses YCbCr. The proposed method is straightforward and miniature yet still gives near state-of-the-art results.