基于多尺度循环网络的运动模糊图像复原方法

Motion blurred image restoration method based on multi⁃scale recurrent network

针对目前图像去模糊恢复细节不好、泛化性能不高的问题,提出了一种基于多尺度循环网络的运动模糊图像复原算法,在编码端将多尺度特征融合模块和残差密集连接模块融合,引入多尺度残差密集型连接模块。同时引入一种注意力机制并将融合了注意力机制的多尺度残差密集连接模块作为网络的基本结构。实验结果表明,与生成对抗网络相比,在GOPRO数据集的PSNR和SSIM最大分别提升了4.13 dB和0.0254 dB,和近年来效果最明显的SRN相比,Kohler数据集上的PSNR和SSIM分别提升了0.31 dB和0.0179 dB,具有更好的泛化性能。

Aiming at the problem that the image deblurring-restoring details is not good and the generalization performance is not high.A motion blur image restoration algorithm based on multi-scale recurrent network is proposed.At the encoder,the Multi-scale Feature Extraction Block and Residual Dense Block are fused,and the Multi-scale Residual Dense Block is introduced.At the same time,an attention mechanism is introduced and the multi-scale residual dense block module integrated with attention mechanism is used as the basic structure of the network,which is beneficial to extract the channel and spatial feature information of the image.The experimental results show that compared with the GAN,the PSNR and SSIM of the GOPRO data set have increased by 4.13 dB and 0.0254 dB,respectively.Compared with the most effective SRN in recent years,the PSNR and SSIM of the Kohler data set have been improved respectively.0.31 dB and 0.0179 dB,with better generalization performance.