基于二阶段特征融合网络的图像超分辨率重建
Super-Resolution Image Reconstruction Based on Feature Fusion Network
杨筝 1程云志 2张文博 3侯彦东 3陈政权2
作者信息
- 1. 黄河水利职业技术学院 电气工程学院,河南开封 475004
- 2. 河南大学计算机与信息工程学院,河南开封 475004
- 3. 河南大学人工智能学院,郑州 450046
- 折叠
摘要
针对现有超分辨率重建算法特征利用率不足和高频细节恢复能力较弱的问题,本文提出了一种网络复杂度较低的二阶段特征融合网络(Two-Stage Feature Fusion Network,TFFN).该网络首先结合稠密连接的方式,通过不同大小的卷积核进行上下采样,提高了网络的特征利用率.然后,设计了二阶段特征融合模块,通过分阶段特征融合的方式,有助于网络专注于高频特征提取.大量实验证明TFFN具有轻量化、高效的特点,与具有代表性的超分辨率重建方法相比,TFFN在重建质量和复杂程度之间取得了较好的平衡.
Abstract
In addressing the challenges posed by insufficient feature utilization and relatively weak high-frequency detail recovery capabilities in existing super-resolution reconstruction algorithms,this paper introduces an innovative two-stage feature fusion network(TFFN)with a relatively low network complexity.The network employs a dense connection approach to utilize convolutional kernels of varying sizes for up and down-sampling,thereby enhancing the comprehensive utilization of features.Subsequently,a two-stage feature fusion module is designed by employing a phased feature fusion approach to help the network focus on the extraction of high-frequency features.Extensive experimentation substantiates the lightweight and efficient characteristics of TFFN.In comparison to representative super-resolution reconstruction methods,TFFN achieves a commendable balance between reconstruction quality and network complexity.
关键词
超分辨率重建/稠密连接/特征融合/神经网络Key words
super-resolution reconstruction/dense connection/feature fusion/neural network引用本文复制引用
基金项目
国家自然科学基金资助项目(61374134)
河南省自然科学基金资助项目(232300421149)
出版年
2024
NETL
NSTL
万方数据