Multi-scale residual attention network for single image dehazing

Abstract

Single image dehazing is a fundamental but difficult task in image processing. Although various deep learning-based methods have achieved good dehazing performance, it is still a challenge to recover a haze-free image without color distortion from a hazy image. The color performance of nowadays dehazing methods needs further improvement. In this paper, the effect of haze on the luminance and chrominance of CIELAB color space is firstly studied. Based on CIELAB color space, a novel multi-scale residual attention network (MSRA-Net) is proposed for single image dehazing. The proposed network consists of two subnetworks: one is designed to remove haze from the luminance channel; another focuses on enhancing the chrominance components and makes the color more realistic. To improve the performance of MSRA-Net, the multi-scale residual attention block (MSRA-block) and feature aggregation building block (FAB-block) are presented. Both the subjective and objective performance show that the proposed method outperforms the existing dehazing algorithms, especially in terms of color performance.