基于改进的残差网络的图像去雨方法

Image Rain Removal Method Based on Improved Residual Network

基于深度卷积神经网络(CNN),提出了一种新的深度网络结构,用于去除图像中的雨痕。受残差网络(ResNet)的启发,通过改变映射形式来简化学习过程,论文提出了一个改进的深度残差网络来直接缩小从输入到输出的映射范围,这使得学习过程更容易。为了进一步改善衰落结果,利用先验图像领域的知识,通过在训练过程中关注高频细节,去除背景干扰,并将模型聚焦在图像中的雨水结构上。虽然是在合成数据上训练网络,但可以发现学习网络能很好地适应现实世界的测试图像。实验表明,论文所提出的方法在定性和定量测量方面都明显优于其他图像去雨的最新方法。

Based on deep convolutional neural network(CNN),a new deep network structure is proposed to remove rain marks from the image.Inspired by the residual network(ResNet),by changing the mapping form to simplify the learning process,an improved depth residual network is proposed to directly reduce the mapping range from input to output,which makes the learning process easier.In order to further improve the fading results,the knowledge of prior image fields is used to focus on the rainwater structure in the image by focusing on high-frequency details during the training process to remove background interference.Al⁃though the network are trained on synthetic data,it can be found that the learning network is well adapted to the real world test imag⁃es.Experiments show that the proposed method is better than the newest methods for rain removal of other images in both qualitative and quantitative measurements.