Single Image Dehazing Based on Convolutional Neural Network Using Boundary Constraint

Abstract

Images captured under hazy weather conditions are often seriously degraded. As a consequence, the captured outdoor images often have details corrupted or lost and fading colors or low contrast. Therefore, image dehazing techniques are highly desired to remove or alleviate such degradation. However, it still remains a challenging task due to its severely ill-posed nature. In this paper, a trainable and efficient single image dehazing method is presented to restore a haze-free image. The proposed method employs the boundary constraint to improve the medium transmission map and estimates the atmosphere light based on quad-tree subdivision to improve the visual quality of the haze-free image. Experiments results on synthetic images and real-world ones confirm the effectiveness of proposed method.