摘要
由于JPEG有损压缩严重影响图像质量,为去除JPEG压缩伪影,提出一种基于卷积神经网络的联合图像滤波器。该文借助可逆下采样层提取特征信息并构建引导特征图;融合输入图与引导特征图的特征,将公共结构映射到目标图像上;利用跳跃连接得到最终滤波输出。为提高泛化能力,该网络基于残差学习联合训练多种压缩水平的图像。相较于几种经典算法,所提算法不仅可以有效去除各种压缩水平的JPEG压缩伪影、提取更清晰的边缘结构,而且具有更高的峰值信噪比和结构相似性,有效改善了图像质量。
Because lossy JPEG compression seriously affects the image quality,a jointly image filters based on convolutional neural network is proposed to reduce jpeg compression artifacts.It extracted the feature information by a reversible down-sampling layer to construct the guided feature image.The features from the input image and guided feature image were fused,which was to optionally transfer the common structure to the target image.The filtering output was obtained by skip connection.Based on residual learning,the network jointly trained samples of various compression levels to improve the generalization ability.Compared with the several classical algorithms,the proposed algorithm not only effectively removes the JPEG compression artifacts at various compression levels and extracts clearer edge structures,but also has higher peak signal to noise rate and structural similarity,and the image quality is improved significantly.
作者
聂栋栋
张琰
程乾坤
赵雪晴
Nie Dongdong;Zhang Yan;Cheng Qiankun;Zhao Xueqing(School of Science,Yanshan University,Qinhuangdao 066000,Hebei,China;JIER Machine-Tool Group Co.,Ltd.,Press and Automation Company,Jinan 250000,Shandong,China)
出处
《计算机应用与软件》
北大核心
2026年第1期233-240,共8页
Computer Applications and Software
基金
国家自然科学基金项目(61807029)
河北省自然科学基金项目(F2019203427)
河北省自然科学基金项目面上项目(A2020203021)。
关键词
JPEG压缩
联合滤波
引导特征图
残差学习
JPEG compression
Joint filtering
Guided feature image
Residual learning
