摘要： 基于卷积神经网络的单图像超分辨率模型网络结构过深，导致高频信息丢失以及模型体积庞大等问题。提出一种由多个残差模块构成的多窗口残差网络优化模型，通过使用多个不同尺寸的窗口对同一特征图进行提取，获取更丰富的高频与低频信息，并过滤出深层网络的所需特征。残差模块中较大尺寸的窗口采用较小尺寸的滤波器和多层映射层叠加组成，可在减少参数总量的同时增强网络的非线性表达能力。实验结果表明，与A+、SRCNN、ESPCN等模型相比，该模型可有效利用原始图像信息生成细节更清晰的超分辨率图像，且在主观视觉效果与客观评价指标上均有所提升。

关键词: 单图像超分辨率重建, 多窗口残差网络, 卷积神经网络, 深度学习, 特征融合

Abstract: The single image super-resolution model based on Convolutional Neural Network (CNN) is faced with problems including the loss of high-frequency information and the large model size caused by the depth of deep network structure. To address the problems, this paper proposes an optimization model of multi-windows residual network composed of multiple residual modules. By using multi-windows with different sizes to extract the same feature map, more abundant high-frequency and low-frequency information can be obtained, and the features required by deep network are filtered out. The larger window in the residual module is composed of the superposed smaller filter and multiple mapping layers, which can reduce the total number of parameters and enhance the nonlinear expression ability of the network. Experimental results show that compared with A+, SRCNN, ESPCN and other models, this model can effectively use the original image information to generate super-resolution images with clear details, and improve the subjective visual effect and objective evaluation index.

Key words: single image super-resolution reconstruction, multi-windows residual network, Convolutional Neural Network (CNN), deep learning, feature integration

中图分类号: 

• TP751.1