High-Resolution Image Dehazing with Respect to Training Losses and Receptive Field Sizes

Haze removal is one of the essential image enhancement processes that makes degraded images visually pleasing. Since haze in images often appears differently depending on the surroundings, haze removal requires the use of spatial information to effectively remove the haze according to the types of image region characteristics. However, in the conventional training, the small-sized training image patches could not provide spatial information to the training networks when they are relatively very small compared to the original training image resolutions. In this paper, we propose a simple but effective network for high-resolution image dehazing using a conditional generative adversarial network (CGAN), which is called DeHazing GAN (DHGAN), where the hazy patches of scale-reduced training input images are applied to the generator network of DHGAN. By doing so, DHGAN can capture more global features of the haziness in the training image patches, thus leading to improved dehazing performance. Also, DHGAN is trained based on the largest binary cross entropy loss among the multiple outputs so that the generator network of DHGAN can favorably be trained in accordance with perceptual quality. From extensive training and test, our proposed DHGAN was ranked in the second place for the NTIRE2018 Image Dehazing Challenge Track2: Outdoor.

1. Introduction

Image enhancement is a crucial process before consuming the degraded images. As damaged images degrade the visual perception of both human and machines, it is important to remove the disturbing part of the images. In many cases, image enhancement is required as a preprocessing. One of these challenging tasks is a haze removal. Haze does not appear consistently across different images and differs from various regions even within a single image. So, it is often non-uniformly distributed, depending on the surrounding atmosphere.