基于LSTM的递归网络图像去雨算法

Recursive network image derainingalgorithm based on LSTM

随着深度学习的发展的热潮,单幅图像去雨得到了很大的发展。然而由于雨图像在方向、大小和雨密度的雨纹的不同,使得去雨的工作变得更困难。针对以上问题,提出了一种基于LSTM的递归图像去雨算法,在特征提取方面采用卷积块和残差块相结合,并运用长短期记忆模块(LSTM)进行多层递归去雨,最后通过注意力融合模块进一步提取雨纹特征,对不同方向、大小等雨纹有较强的学习能力,较好地保留了图像的细节,通过在真实数据集和合成数据集上进行实验,证明了该方法的有效性,通过与其他算法的比较,在客观指标和主观效果上优于它们。主观效果去雨更彻底,图像细节更加清晰。在合成数据集Rain100H上PSNR达到30.48,SSIM为0.91,在Rain100L上PSNR达到38.05,SSIM为0.98。

With the boom in the development of deep learning,single image de-rain has been greatly developed.However,due to the difference in the direction,size and density of rain in the rain image,the work of de-raining becomes more difficult.Aiming at the above problems,a recursive image de-raining algorithm based on LSTM is proposed,which combines convolutional blocks and residual blocks in feature extraction,and uses long-term short-term memory module(LSTM)for multi-layer recursive rain removal,and finally further extracts rain pattern characteristics through attention fusion module,has strong learning ability for rain patterns in different directions,sizes,etc.,and better retains the details of the image,and proves the effectiveness of the method by experimenting on the real data set and synthetic data set.By comparing with other algorithms,they are superior to them in objective indicators and subjective effects.The subjective effect goes rain more thoroughly,and the image details are clearer.PSNR reached 30.48 and SSIM 0.91 on the synthetic dataset Rain100 H,PSNR reached 38.05 and SSIM reached 0.98 on Rain100 L.