基于加速残差网络的图像超分辨率重建方法

针对深层网络架构的图像超分辨率重建任务中存在网络参数多、计算复杂度高等问题，提出了一种基于加速残差网络的图像超分辨率重建方法。首先，构建一个残差网络对低分辨率图像与高分辨率图像之间的高频残差信息进行重建，以减少冗余信息的深层网络传输过程，提高重建效率；然后，通过特征收缩层对提取的低分辨率特征图进行降维，从而以较少的网络参数实现快速映射；之后，对高分辨率特征图通过特征扩展层进行升维，从而以较丰富的信息重建高频残差信息；最后，将残差与低分辨率图像求和得到重建的高分辨率图像。实验结果表明，该方法取得的峰值信噪比（PSNR）及结构相似性（SSIM）均值结果较基于卷积神经网络的图像超分辨率（SRCNN）取得的结果分别提升了0.57 dB和0.013 3，较基于中间层监督卷积神经网络的图像超分辨率重建（ISCNN）取得的结果分别提升了0.45 dB和0.006 7；在重建速度方面，以数据集Urban100为例，较现有方法提高了1.5~42倍。此外，将该方法应用于运动模糊图像的超分辨率重建时，获得了优于超深卷积神经网络的图像超分辨率（VDSR）的性能。所提方法以较少的网络参数快速获得较好的重建质量，并且为图像超分辨率重建提供了新的思路。

Image super-resolution reconstruction method based on accelerated residual network

To solve the problems of multiple network parameters and high computational complexity in image super-resolution reconstruction of deep network architecture, an image super-resolution reconstruction method based on accelerated residual network was proposed. Firstly, a residual network was constructed to reconstruct the high-frequency residual information between low-resolution image and high-resolution image, so as to reduce the deep network transmission process of redundant information and improve the reconstruction efficiency. Secondly, the dimensionality of the extracted low-resolution feature map was reduced by the feature shrinking layer to realize fast mapping with fewer network parameters. Thirdly, the dimensionality of the high-resolution feature map was increased by the feature expanding layer to reconstruct the high-frequency residual information with the rich information. Finally, the residual and low-resolution images were summed to obtain the reconstructed high-resolution image. Experimental results show that the Peak Signal-to-Noise Ratio (PSNR) and the Structural SIMilarity (SSIM) mean results obtained by the proposed method are 0.57 dB and 0.013 3 higher than those obtained by Super-Resolution using Convolutional Neural Network (SRCNN) respectively, and 0.45 dB and 0.006 7 higher than those obtained by Intermediate Supervision Convolutional Neural Network (ISCNN). In terms of reconstruction speed, using dataset Urban100 as example, the proposed method is 1.5 to 42 times faster than the existing methods. In addition, when this method is applied to the super-resolution reconstruction of motion blur images, it has the performance better than image Super-Resolution using Very Deep convolutional network (VDSR). The proposed method achieves better reconstruction quality with fewer network parameters and provides a new idea for image super-resolution reconstruction.