基于高效多尺度视觉的图像去雨网络  

Image Rain Removal Network Based on Efficient Multi-scale Vision

作　　者：巩珂菘 Gong Kesong(College of Transportation,Shandong University of Science and Technology,Qingdao,China)

机构地区：[1]山东科技大学交通学院,山东青岛

出　　处：《科学技术创新》2025年第8期21-24,共4页Scientific and Technological Innovation

摘　　要：为解决雨水遮挡降低图像质量,影响下游视觉任务的准确性,现有方法多局限于去除特定类型的雨条的问题,本文提出了一种基于高效多尺度视觉注意力机制的多阶段视觉Transformer模型,通过多尺度注意力模块及Transformer结构,有效地去除不同尺度的雨痕,并保留图像细节。实验在Rain12600和Rain1400数据集上验证了该模型的优越性。该研究为图像去雨技术提供了新的思路,未来可进一步优化模型以应对其他恶劣天气。In order to solve the problem that rain shielding reduces image quality and affects the accuracy of downstream visual tasks,existing methods are mostly limited to removing certain types of rain strips.This paper proposes a multi-stage vision Transformer model based on an efficient multi-scale visual attention mechanism,which can effectively remove rain marks of different scales and retain image details.Experiments on Rain12600 and Rain1400 data sets demonstrate the superiority of the proposed model.This study provides new ideas for image rain removal technology,which can be further optimized to cope with other severe weather in the future.

关 键 词：深度学习 图像去雨 Transformer模块 高效多尺度注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]