A Deep Meta-Learning Neural Network for Single Image Rain Removal
Yihong Lu, Jianyong Cai, Hua Zheng, Yuanqiang Zeng · 2020
Due to various interference factors, image degradation seriously retards the precision and efficiency of target tracking and recognition. Consequently, image restoration becomes a significant issue in the field of computer vision. In this paper, a novel de-rain network, called Meta-DerainNet was proposed to remove rain streaks in various conditions. With the strongly generation ability, Model-Agnostic Meta-Learning (MAML) has been proven to be able to accomplish the task of removing rain streaks from single image. However, in order to combine the MAML algorithm with rain removal tasks, there are many parts need to be improved. The datasets with specific sample methods were designed to train the generalization ability of the network, and the network structured of MAML was improved for more efficiently extract feature map. The modified network was successfully applied to remove rain streaks under various conditions. Experimental results on real-world data show that the method proposed in this paper performs better than other advanced methods on PSNR and SSIM. Furthermore, our method can also be applied to process other image interference removal factors, such as medical image recognition, self-driving and other fields.