Raindrop Removal via Median-Prior-Guided Gating Mechanism

Mengmeng Zhang, Bowen Shao, Qiyu Rong, Hui Zhang, Yi Ren, Hongyuan Jing, Tianxu Cui, Huihui Bai, Zhi Liu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026, pp. 2802-2811

In real-world rainy scenarios, raindrop morphologies exhibit significant diversity. Among them, long streak-like raindrops pose a greater challenge to removal tasks due to their large-area occlusion of pixel information. Existing deraining methods often generate noticeable visual artifacts when handling such raindrops, severely compromising the quality and fidelity of the reconstructed images. To address the aforementioned issues, this paper proposes a Median-Prior-Guided Gating Mechanism (MPGGM). This mechanism first applies median filtering to multi-frame sequential images along the temporal dimension to obtain a median prior image that effectively suppresses transient raindrop noise. Subsequently, an adaptive pixel-level fusion of the median prior image and the original input image is performed through a spatial gating module. Specifically, the mechanism utilizes the background pixel information carried by the median prior image to effectively compensate for and restore the degraded regions occluded by raindrops in the original input image. Meanwhile, with the help of a spatial weight map, the contribution of the median prior image in non-raindrop regions is explicitly suppressed, thereby achieving a fine-grained adaptive balance between the effective restoration of raindrop regions and the background fidelity of non-raindrop regions. Experimental results demonstrate the effectiveness of our method, which achieved third place in the NTIRE 2026 Challenge on Day and Night Raindrop Removal for Dual-Focused Images.