G3RHW: A Unified Model for the Generation and Removal of Real Rain and Haze Weather

• Published in: 2023 IEEE 7th Information Technology and Mechatronics Engineering Conference (ITOEC) Vol. 7; pp. 491 - 498
• Main Authors: Zhao, Qing, Luo, Yu, Hao, Cha
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.09.2023
• Subjects: Degradation; GAN; haze removal; high-order degradation; image synthesis; Mechatronics; Rain; rain removal; Redundancy; Supervised learning; Training; Training data
• ISSN: 2693-289X
• DOI: 10.1109/ITOEC57671.2023.10291238

Supervised deep learning is widely used in the field of rain and haze weather removal currently and has achieved excellent results. However, it is challenging to learn the potential relationship between rain and haze image input and clean image output in learning based network training because it is extremely difficult to obtain the matched Noising-Clean image data pairs in the real world for supervised learning model train. To overcome this obstacle, in this paper, we propose a high-order rain and haze degradation process to generate a large number of more realistic and complex rain haze image training pairs, and build an end-to-end unified model G3RHW for rain haze generation and removal based on generative adversarial networks. Extensive experiments demonstrate that the proposed method achieves more superior image detail recovery and the removal of the effect of rain and haze results on both synthetic and realistic dataset than state-of-the-art methods fine-tuned for adverse weather removal.