基于注意力编解码器及多残差网络的逆半色调方法
An inverse halftoning method based on encoder-decoder with attention and multi-residual network
邬凡 1杨俊 2桂江生3
作者信息
- 1. 浙江理工大学计算机科学与技术学院,杭州 310018;嘉兴学院信息科学与工程学院,浙江嘉兴 314001
- 2. 嘉兴学院信息科学与工程学院,浙江嘉兴 314001
- 3. 浙江理工大学计算机科学与技术学院,杭州 310018
- 折叠
摘要
针对当前逆半色调方法恢复的图像存在细节不清晰甚至丢失的问题,提出了一种基于注意力编解码器及多残差网络(Encoder-decoder with attention and multi-residual network,EDAMRNet)的逆半色调方法.首先,设计融合注意力机制的编解码器结构,在其跳跃连接处添加非对称特征融合模块,以有效提取图像上下文信息;然后,构造多残差网络,捕获并保留图像空间细节信息;最后,应用监督注意力模块对图像上下文信息进行加强,再传递到多残差网络,以恢复出高质量的连续色调图像.实验结果表明:该方法与现有最优方法相比,在Urban100和Manga109数据集下的峰值信噪比平均值均提高了 0.1 dB,结构相似性平均值分别提高了 0.0010和0.0005.该方法能够在提取图像上下文信息的同时保留图像空间细节信息,可更好地恢复图像纹理信息,提高图像清晰度,为图像逆半色调方法研究提供了一种新的方案.
Abstract
An inverse halftoning method based on encoder-decoder with attention and multi-residual network(EDAMRNet)was proposed to address the issue of unclear or even lost details in image restoration using current reverse halftone methods.Firstly,the encoder-decoder structure with attention was designed,and the asymmetric feature fusion modules were added at its skip connections to effectively capture image context information.Then,the multi-residual network was constructed to capture and retain spatial details of the image.Finally,the supervised attention module was applied to enhance the image context information,which was then transmitted to multi-residual network to restore high-quality continuous-tone images.The experimental results showed that compared to the existing optimal method,the proposed method improved the average peak signal-to-noise ratio by 0.1 dB and the average structural similarity by 0.0010 and 0.0005,respectively on the Urban100 and Manga109 datasets.The method can extract image context information while preserving spatial details,better restoring image texture information and improving image clarity.It provides a new scheme for the study of image inverse halftoning.
关键词
逆半色调/图像恢复/注意力机制/编解码器/多残差网络/清晰度Key words
inverse halftoning/image restoration/attention mechanism/encoder-decoder/multi-residual network/clarity引用本文复制引用
基金项目
浙江省基础公益项目(LGG22F020021)
嘉兴市科技计划(2021AY10071)
出版年
2024
NETL
NSTL
万方数据