一种基于深度学习的图像超分辨率重构方法

A Super-resolution Image Reconstruction Method Based on Deep Learning

近年来,基于卷积神经网络(CNN)的单幅图像超分辨率重构得到了广泛应用.然而,随着网络不断加深,也同时出现了参数过多、计算代价过大和难以训练等问题.为解决上述问题,提出一种新的深度残差密集网络(DRDN)框架并应用于单幅图像超分辨率重建.首先,网络通过密集连接充分利用了低分辨率图像从浅层到深层的各层特征,为超分辨率重构提供更多的低分辨率图像信息;其次,为了充分融合全局特征信息,通过残差学习的方式进行融合重构,同时为了缓解深层网络带来的训练困难等问题,网络采用多路跳步连接,使误差更加快速地传到各层网络;最后,将该方法与深度递归残差网络(DRRN)方法在公共数据集上进行了实验比较,结果表明DRDN在网络稳定性、时间效率、收敛速度和重建效果等方面都优于DRRN.

In recent years,single image super-resolution reconstruction based on convolutional neural network(CNN)has been widely used.However,with the deepening of the network,problems such as too many parameters,too much computational cost,and difficulty in training are also occurred.To solve the above problems,a new deep residual dense network(DRDN)framework is proposed and applied to single image super-resolution reconstruction.Firstly,the network makes full use of the features of low-resolution images from shallow to deep layers by using dense connections to provide more low-resolution image information for super-resolution reconstruction.Secondly,in order to fully integrate the global feature information,the fusion and reconstruction are carried out by residual learning.At the same time,in order to alleviate the training difficulties caused by the deep network,the network uses multi-hop connection,the errors can be transmitted to all layers of the network more quickly.Finally,an experimental comparison is made between this method and DRRN method on a common data set.The results show that the DRDN is superior to the DRRN in terms of network stability,time efficiency,convergence speed,and reconstruction effect.