基于双流U型的单图像超分辨率重建方法研究
Research on Single Image Super-Resolution Reconstruction Method Based on Dual-Stream U-Shape
李冬 1杨思路 2张恒 1王晓明1
作者信息
- 1. 西华大学计算机与软件工程学院,四川成都 610039
- 2. 成都东软学院高等职业技术学院,四川成都 611844
- 折叠
摘要
深度神经网络在图像超分辨方面吸引了广泛的关注.然而,许多超分辨网络忽略了多种不同特征信息中对图像重建的贡献.针对以上的问题,设计了一种双流U型超分辨网络,该网络利用U型结构中所产生的不同尺度的特征图来重建图像特征,并且还使用新颖的投影结构来增强差异特征的表达.此外为了嵌入到现实中的小型设备,还设计了一种轻量型网络.实验表明,设计的网络在流行的数据集上取得了不错的指标值和高质量的视觉效果.
Abstract
Deep neural networks have attracted widespread attention in image super-resolution.However,many super-resolution networks ignore the contribution of multiple different feature information to image recon-struction.To address the above problems,a double-stream U-shaped super-resolution network is designed,which utilizes the different scales of feature maps generated in the U-shaped structure to reconstruct image fea-tures,and also uses a novel projection structure to enhance the representation of disparate features.In addition,a lightweight network is designed in order to be embedded into realistic small devices.Experiments show that the designed networks achieve good metric values and high quality visualization on popular datasets.
关键词
深度神经网络/超分辨/特征提取流/轻量型网络Key words
deep neural network/super-resolution/feature extraction flow/lightweight network引用本文复制引用
基金项目
四川省自然科学基金(2022NSFSC0533)
出版年
2024
NETL
NSTL
万方数据