Abstract:

Super-resolution reconstruetion methods based on Convolutional Neural Network（CNN） are confronting the problems of small receptive field, single feature extraction scale and disappearing of gradient information. In order to solve these problems, a single image super-resolution reconstruction method is proposed based on multi-scale residual network. By using multi-scale feature extraction and feature information fusion, the problem of insufficient image feature extraction is solved. The combination of local residual learning and global residual learning improves the efficiency of information flow and greatly reduces the phenomenon of gradient disappearance. Experiments are carried out on common test sets such as Set5, Set14 and BSD100, the experimental results of the proposed method are higher than the results of five comparison methods. Compared with the SRCNN, the average PSNR is improved by 0.74 dB and the average SSIM is improved by 0.014 3 dB. Compared with the VDSR, the average PSNR is improved by 0.12 dB and the average SSIM is improved by 0.002 5 dB.

Key words: super-resolution reconstruction, Convolution Neural Network（CNN）, residual learning, multi-scale feature

摘要：

针对目前提高图像分辨率的卷积神经网络存在的特征提取尺度单一以及梯度消失等问题，提出了多尺度残差网络的单幅图像超分辨率重建方法。采用多尺度特征提取和特征信息融合，解决了对图像细节特征提取不够充分的问题；将局部残差学习和全局残差学习相结合，提高了卷积神经网络信息流传播的效率，减轻了梯度消失现象。在Set5、Set14和BSD100等常用测试集上进行了实验，该方法的实验结果均优于其他5种方法，相比于SRCNN方法，平均PSNR提升了0.74 dB，平均SSIM提升了0.014 3 dB；相比于VDSR方法，平均PSNR提升了0.12 dB，平均SSIM提升了0.002 5 dB。

关键词: 图像超分辨率重建, 卷积神经网络, 残差学习, 多尺度特征