一种多尺度循环残差注意的单幅图像去雨方法

A MULTI-SCALE RECURRENT RESIDUAL ATTENTION FOR SINGLE IMAGE DERAINING

  • 摘要: 目前基于卷积神经网络的去雨方法,存在雨纹残留、图像模糊等问题。为此提出一种基于多尺度特征提取和循环残差注意的单幅图像去雨方法。通过构建多尺度拉普拉斯金字塔得到多尺度特征图,再设计循环残差注意模块加强阶段间联系、提取深度特征、增强重要特征权重,更好地去除雨纹并保留了图像细节。实验结果表明,该方法的去雨效果优于其他去雨算法。

     

    Abstract: At present, the rain removal methods based on convolution neural network still suffer from residual rain streaks and details lost. This paper proposes a novel network for single image deraining including the multi-scale feature extraction module and the recurrent residual attention. The multi-scale feature map was obtained by constructing the multi-scale Laplacian pyramid. The recurrent residual attention module was designed to promote the connection between stages, extract depth features and enhance the weight of important features, so as to better remove rain streak and preserve the image details. Experimental results demonstrate that the proposed method performs favorably against other state-of-the-art methods.

     

/

返回文章
返回