基于高效通道注意力的多阶段图像去雨网络

A Multi-stage Image Deraining Network Based on Highly Efficient Channel Attention

针对现有图像去雨算法不能更好地保留图像背景细节的问题,提出一种基于高效通道注意力的多阶段图像去雨网络。首先,网络使用3×3卷积提取雨图的浅层特征并传递给高效通道注意力模块,为不同的特征通道分配不同的权重;然后,传递给3个并行阶段,在前2个阶段中,使用编码-解码器进行多尺度特征提取,减少雨纹信息丢失,其中使用Transformer模块抑制无用信息传递;最后,在第3个阶段使用初始分辨率模块代替编码-解码器,从而保留输出图像的精细特征。实验结果表明,所提算法在Rain800、Rain12、Rain100L和Rain100H公开测试集上的结构相似性分别为0.830、0.968、0.960和0.944,峰值信噪比分别为27.33 dB、35.27 dB、36.79 dB和28.94 dB。所提算法相比于经典和新颖的图像去雨算法,在去除雨纹和恢复背景细节上具有更好的效果。

To address the problem that existing image deraining algorithms cannot well preserve the image background details a multi-stage image deraining network based on highly efficient channel attention is proposed.Firstly the network uses 3×3 convolution to extract shallow features of the rain map and passes them on to the Highly Efficient Channel Attention Block(HECAB)assigning different weights to different feature channels.Then it is transfered to three parallel stages.In the first two stages the encoder-decoder is used for multi-scale feature extraction to reduce rain pattern information loss where the Transformer block is used to suppress useless information transfer.Finally in the third stage the initial resolution block is used to replace the encoder-decoder thus preserving the fine features of the output image.The experimental results show that:1)The structural similarities of the proposed algorithm on the public test sets of Rain800 Rain12 Rain100L and Rain100H are 0.8300.9680.960 and 0.944 and the peak signal-to-noise ratios are 27.33 dB 35.27 dB 36.79 dB and 28.94 dB;and 2)Compared with classical and novel image deraining algorithms the proposed algorithm has better results in removing rain patterns and recovering background details.