摘要
图像超分辨率旨在通过软件技术从低分辨率图像中获得高分辨率图像。受深层神经网络对非线性关系强大表示启发,提出一种基于多尺度密集连接网络的图像超分辨算法。利用多尺度和密集连接思想设计了两个并行子网络提取图像特征,一个子网络中引入多尺度卷积层以提取低分辨图像的多种特征,另一个子网络则利用密集连接模块加深网络结构尽可能提取丰富的纹理特征,同时还可以避免模型训练过程中梯度消失的问题。最后对两个子网络提取的特征求残差并对其行重构得到高分辨率图像。仿真结果表明,提出的算法无论在客观评价还是视觉效果上均优于其它同类超分辨算法。
Image super-resolution aims to obtain a high-resolution image from a single low-resolution image through software technology. Inspired by the powerful representation of nonlinear relationships by deep neural networks, this paper proposes an image super-resolution algorithm based on multi-scale densely connected networks. Using the idea of multi-scale and dense connection, two parallel sub-networks are designed to extract image features. One sub-network uses a multi-scale convolution layer to extract various features of low-resolution images, and the other sub-network uses dense connections. The module can avoid the problem of gradient disappearance during model training. Finally, the residuals of the features extracted by the two sub-networks are calculated and reconstructed. Both the quantitative assessment results and the visual assessment confirm that the proposed network yielded high-resolution images, which are superior to the ones in comparison with the state-of-the-art methods.
作者
贺伟
王丹阳
HE Wen;WANG Dan-yang(Faculty of Elctrical and Control Engineering,Henan University of Urban Construction,Pingdingshan Henan,467306,China)
出处
《计算机仿真》
北大核心
2021年第11期259-265,共7页
Computer Simulation
基金
国家自然科学基金资助项目(61803145)
河南省高等学校重点科研项目(21B510002)。
关键词
深度学习
卷积神经网络
多尺度密集连接网络
图像超分辨
Deep learning
Convolutional neural network
Multi-scale densely connected network
Super-resolution
