Single Image Dehazing via Transmission Map Estimation using Deep Neural Networks

Abstract

Myriads of existing approaches which perform single image dehazing make use of the widely adopted Atmospheric Scattering Model. De-hazing though this model involves the estimation of 2 key components - the transmission of haze through the scene and global atmospheric light. Depending on various factors, such as time of day, photography equipment, etc., the aforementioned components of any particular image can vary wildly and can prove quite difficult to correctly estimate. Some previous approaches that exist make use of either premeditated priors to estimate these components or end-to-end neural networks to fully reconstruct a de-hazed image. This paper delves into the use of a deep convolutional, U-Net based segmentation network to obtain the medium transmission map from the input hazy image is explored. The model is trained using a supervised approach with a samples of hazy scenes and their corresponding medium transmissions. The estimation of global atmospheric light of a scene is done using a modification of the dark channel prior method, making use of the Y(luma) component of the YUV representation of the hazy image. Experimental and benchmarking results on 3 different testing datasets are presented, which show that this system can produce high quality haze-free images and can do so efficiently and reliably.