计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (13): 1-7.DOI: 10.3778/j.issn.1002-8331.1812-0315

• 热点与综述 • 上一篇    下一篇

改进的卷积神经网络单幅图像超分辨率重建

曾接贤,倪申龙   

  1. 南昌航空大学 江西省图像处理与模式识别重点实验室,南昌 330063
  • 出版日期:2019-07-01 发布日期:2019-07-01

Improved Super-Resolution Reconstruction of Single Image Based on Convolution Neural Network

ZENG Jiexian, NI Shenlong   

  1. Key Laboratory of Jiangxi Province for Image Processing and Pattern Recognition, Nanchang Hangkong University, Nanchang 330063, China
  • Online:2019-07-01 Published:2019-07-01

摘要: 针对经典的基于卷积神经网络的单幅图像超分辨率重建方法网络较浅、提取的特征少、重建图像模糊等问题,提出了一种改进的卷积神经网络的单幅图像超分辨率重建方法,设计了由密集残差网络和反卷积网络组成的新型深度卷积神经网络结构。原始低分辨率图像输入网络,利用密集残差学习网络获取更丰富的有效特征并加快特征梯度流动,其次通过反卷积层将图像特征上采样到目标图像大小,再利用密集残差学习高维特征,最后融合不同卷积核提取的特征得到最终的重建图像。在Set5和Set14数据集上进行了实验,并和Bicubic、K-SVD、SelfEx、SRCNN等经典重建方法进行了对比,重建出的图像在整体清晰度和边缘锐度方面更好,另外峰值信噪比(PSNR)平均分别提高了2.69?dB、1.68?dB、0.74?dB和0.61?dB。实验结果表明,该方法能够获取更丰富的细节信息,得到更好的视觉效果,达到了图像超分辨率的增强任务。

关键词: 图像超分辨率重建, 深度学习, 卷积神经网络, 密集残差学习, 反卷积

Abstract: Aiming at the problems of classical super-resolution reconstruction method based on convolutional neural network, such as shallow network, less extracted features and blurred image reconstruction, an improved single image super-resolution based on convolutional neural network is proposed. In this way, a novel deep convolutional neural network structure consisting of a dense residual network and a deconvolution network is designed. Firstly, the original low-resolution image input network uses the dense residual learning network to obtain richer effective features and accelerate the feature gradient flow. Secondly, the image features are up sampled to the target image size through the deconvolution layer, and then the dense residual learning is used to get high dimensional features. Finally, the features extracted by different convolution kernels are obtained to obtain the final reconstructed image. Experiments are performed on the Set5 and Set14 datasets and compared with classical reconstruction methods such as Bicubic, K-SVD, SelfEx and SRCNN. The reconstructed images are better in terms of overall sharpness and edge sharpness. In addition, the Peak Signal-to-Noise Ratio(PSNR) averaged by 2.69 dB, 1.68 dB, 0.74 dB and 0.61 dB, respectively. The experimental results show that the proposed method can obtain more detailed information, which get better visual effects and achieve the enhanced task of image super-resolution.

Key words: image super-resolution reconstruction, deep learning, convolutional neural network, dense residual learning, deconvolution