基于密集残差网络和注意力机制的图像超分辨研究
Research on image super-resolution based on dense residual network and attention mechanism
俞成海 1胡异 1卢智龙 1叶泽支1
作者信息
- 1. 浙江理工大学计算机科学与技术学院,浙江 杭州 310018
- 折叠
摘要
针对现有的图像超分辨重建算法特征信息提取不充分的问题,基于SRResNet[1]网络的生成器部分,引入混合注意力模块和密集残差模块,以提取图像的多尺度特征.混合注意力模块集成通道注意力和自注意力机制,可以聚焦关键特征;密集残差模块通过堆积多个残差密集块学习多级特征,并采用改进的密集连接方式提高特征复用效率.模型在各基准数据集上对比当前的优秀重建算法有0.1～1db的提升,为单图像超分辨率任务提供了有效的方案.
Abstract
To address the problem of insufficient feature information extraction in existing image super-resolution reconstruction algorithms,the hybrid attention modules and dense residual modules are introduced into the generator part of the SRResNet network to extract multi-scale features of images.The hybrid attention module integrates channel attention and self-attention mechanisms to focus on critical features.The dense residual module learns multi-level features by stacking multiple dense residual blocks and adopts improved dense connection method to improve feature reuse efficiency.The model achieves 0.1-1db improvement over current excellent reconstruction algorithms on various benchmark datasets,providing an effective solution for single image super-resolution tasks.
关键词
密集残差网络/注意力机制/图像超分辨重建/改进密集连接Key words
dense residual network/attention mechanism/image super-resolution reconstruction/improved dense connection引用本文复制引用
出版年
2023
NETL
NSTL
维普
万方数据