This paper addresses the challenging problem of rain removal from single images, even in heavy rain conditions with significant rain streak accumulation. The authors propose a novel deep learning architecture that includes a new rain image model and a multi-task deep network. The model incorporates a binary map to locate rain streak regions, and it can handle both individual rain streaks and overlapping streaks, simulating heavy rain conditions. The multi-task network jointly detects and removes rain, using the binary map to constrain the removal process and preserve background details. Additionally, a contextualized dilated network is introduced to enhance contextual information, improving the representation for rain detection. The recurrent rain detection and removal network iteratively removes rain streaks, effectively handling complex rain scenarios. Extensive experiments on real images, particularly in heavy rain conditions, demonstrate the effectiveness of the proposed method, outperforming state-of-the-art methods. The authors also provide publicly available codes and datasets.This paper addresses the challenging problem of rain removal from single images, even in heavy rain conditions with significant rain streak accumulation. The authors propose a novel deep learning architecture that includes a new rain image model and a multi-task deep network. The model incorporates a binary map to locate rain streak regions, and it can handle both individual rain streaks and overlapping streaks, simulating heavy rain conditions. The multi-task network jointly detects and removes rain, using the binary map to constrain the removal process and preserve background details. Additionally, a contextualized dilated network is introduced to enhance contextual information, improving the representation for rain detection. The recurrent rain detection and removal network iteratively removes rain streaks, effectively handling complex rain scenarios. Extensive experiments on real images, particularly in heavy rain conditions, demonstrate the effectiveness of the proposed method, outperforming state-of-the-art methods. The authors also provide publicly available codes and datasets.