Deep multi-scale network for single image dehazing with self-guided maps

• Published in: Signal, image and video processing image and video processing, 2023-09, Vol.17 (6), p.2867-2875
• Main Authors: Liu, Jianlei, Yu, Hao, Zhang, Zhongzheng, Chen, Chen, Hou, Qianwen
• Format: Article
• Language: English
• Subjects: Algorithms; Coders; Computer Imaging; Computer Science; Encoders-Decoders; Feature extraction; Image Processing and Computer Vision; Microprocessors; Modules; Multimedia Information Systems; Original Paper; Pattern Recognition and Graphics; Qualitative analysis; Signal,Image and Speech Processing; Vision
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-023-02505-2

A self-guided map, which is obtained from an input hazy image, is useful information as haze removal guidance. The existing end-to-end multi-scale networks tend to recover under-dehazed results due to the lack of a self-guided map. To solve this problem, we propose a deep multi-scale network with self-guided maps for image dehazing, which consists of a pre-processor module and a deep multi-scale network (DMSN). The pre-processor module consists of a pre-dehazer based on the dark channel prior and a pre-dehazer based on gamma-correction, which can generate effective self-guided maps. We concatenate self-guided maps and the hazy image as the DMSN input. Based on the encoder-decoder structure, the DMSN improves feature representation by a new feature extraction block on each scale. The proposed method is experimentally evaluated in detail, and qualitative as well as quantitative analyses are performed. The experimental results show that the proposed algorithm performs favorable against state-of-the-art methods on the widely used dehazing benchmark datasets as well as real-world hazy images.