多尺度沙漏结构的单幅图像去雨算法研究

雨天环境会造成图像模糊、变形,大幅降低图像质量,对于后续的图像分析和应用造成严重影响.单幅图像的去雨算法研究成为热点,然而现有算法存在过度平滑、颜色失真和复杂雨水图像复原能力差等诸多问题,去雨问题难以有效解决.本文提出一种新颖的多尺度沙漏结构的单幅图像去雨算法.首先,针对雨的特征复杂多样的特点,采用多尺度沙漏网络结构,提取并融合多尺度的雨线特征;其次,在沙漏网络内部,引入残差密集模块,使特征在不同级别网络中实现传递和复用,最大限度的提取细节特征和增强网络表达能力;最后,针对雨水不均匀分布的特点,在残差密集网络基础上加入注意模块,提高算法在空间和通道方面特征提取能力,能够处理复杂的雨天图像.实验结果表明,本方法相较于现有算法,能够更好的去除雨线,并且能够最大程度的保留图像细节和颜色信息.

Multi-scale Hourglass Network for Single Image Deraining

The image collected under the rainy weather often suffer from quality degradation,which cause the blur,distortion and some other questions and severely effect the following image analysis and practical application.Single image deraining is an urgent and hot topic,while the existing algorithms have several limitations as over smoothing,color distortion and poor restoration capability in the various rainy scene.To handle these issues,we propose a multi-scale hourglass network to deal with the image deraining problem.Firstly,the structure of multi-scale Hourglass is design to extract and fusion the rain feature,so that multi-scale feature can be used to deal with the complex rainy condition.Further,the residual dense module is involved in the hourglass network to make the feature reuse and transportation smoothly,which can better exact the detail information and network representation capability.Finally,attention mechanism is employed to enhance the residual dense module to increase the feature exaction of spatial and channel aspects,therefore the methods can deal with complex rainy image,especially in the non-uniform distribution.The quantitative and qualitative analysis reveal that the designed network performs better than the recent comparing methods on the deraining performance,as well as the image details and color reservation.