连分式插值结合卷积神经网络的超分辨率重建

Super-resolution reconstruction using deep convolutional neural network combined with continued fraction interpolation

基于卷积神经网络的图像超分辨率重建算法是数字图像处理领域近年来的研究热点。针对低分辨率图像在预处理时使用双三次插值导致图像丢失一些重要的高频纹理细节以及网络模型优化问题,文章提出了连分式插值结合卷积神经网络的超分辨率重建方法。在原有的轻量级基于卷积神经网络的超分辨率重建算法(super-resolution convolutional neural net work,SRCNN)网络模型基础上,首先采用Newton-Thiele型连分式插值函数将低分辨率图像插值到目标尺寸;然后利用3个卷积层进行图像特征提取、非线性映射、重建与优化;该文在网络收敛时利用Radam优化算法自适应地调整梯度,并且采用余弦衰减法逐渐降低学习率。实验结果表明,该网络模型能够在轻量级的卷积神经网络下获得更丰富的纹理细节和更清晰的图像边缘。

Image super-resolution reconstruction using deep convolutional neural network is a research hotspot in the field of digital image processing in recent years.In order to solve the problems of the loss of some important high frequency texture details during the preprocessing of low-resolution images by using bicubic interpolation and the optimization of the network model,a super-resolution reconstruction method combining continued fraction interpolation with convolutional neural network was proposed.On the basis of the original lightweight super-resolution convolutional neural net work,SRCNN network model,the Newton-Thiele continued fraction interpolation function was firstly used to interpolate the low-resolution image to the target size.Then three convolution layers were used for image feature extraction,nonlinear mapping,reconstruction and optimization.At the same time,Radam algorithm was used to adjust the gradient adaptively when the network converged,and the cosine attenuation method was used to gradually reduce the learning rate.The experimental results show that the network model designed in this paper can obtain richer texture details and clearer image edges under the lightweight convolutional neural network.