基于双分支融合网络的图像超分辨率重建与增强
Image Super Division Reconstruction and Enhancement Based on Dual Branch Fusion Network
贾世杰 1杨真杰 1孙万鑫1
作者信息
- 1. 大连交通大学 自动化与电气工程学院,辽宁 大连 116028
- 折叠
摘要
针对现有的图像超分算法难以从模糊的低分辨率图像中重建清晰的高分辨率图像的问题,提出了双分支融合网络,通过双分支结构来联合处理图像去模糊增强、图像超分任务.网络整体分为特征提取、特征融合、重建3个阶段.在特征提取阶段,通过以ResNet为基本模式所构建的轻量化残差组、增强稠密残差块来强化对去模糊局部特征、多尺度高频特征的提取;同时为了提升关键区域的特征表达,引入监督注意力模块将提取到的特征进行校准与细化.在特征融合阶段,以像素相乘、通道相加的方式进行融合.在重建阶段,通过多个卷积操作提升空间分辨率.试验结果表明,对于4倍重建任务,输出图像的峰值信噪比(PSNR)在LR-GOPRO、Set5数据集上比GFN网络分别提高了1.34、1.36 dB,且模型的参数减少约50%.
Abstract
Existing image super-resolution methods are difficult to reconstruct clear high-resolution images from blurred low-resolution images.To address the above problems,Dual Branch Fusion Network (DBF-Net)is proposed to jointly handle image deblurring enhancement and image super-segmentation tasks through a dual-branch structure.The network as a whole is divided into three stages:feature extraction,feature fu-sion,and reconstruction.In the feature extraction stage,the light-weighted residual groups and enhanced dense residual blocks built with ResNet as the basic model are used to enhance the extraction of deblurred lo-cal features and multi-scale high-frequency features.Meanwhile,the supervised attention module is intro-duced to calibrate and refine the extracted features in order to improve the feature representation in key re-gions.The feature fusion stage is performed by pixel multiplication and channel adding.The reconstruction stage enhances the resolution of space by multiple convolution operations.Experiment results show that for quadruple reconstruction tasks on LR-GOPRO and Set5 datasets,the Peak Signal-to-Noise Ratio (PSNR)of the output image is reduced by 1.34 dB and 1.36 dB compared to the GFN network,respectively.For the in-crease,the number of model parameters is reduced about 50％.
关键词
超分辨率重建/卷积/双分支/特征融合/稠密残差Key words
super-resolution reconstruction/convolution/two branches/feature fusion/dense residual引用本文复制引用
基金项目
辽宁省教育厅科学研究计划项目(LJKMZ20220826)
出版年
2024
NETL
NSTL
万方数据