基于拉普拉斯金字塔残差网络的多尺度图像压缩研究  

Research on Multiscale Image Compression Based on Laplacian Pyramid Residual Network

作　　者：田学军[1] 章文强 马梓轩 陈良哲 叶卉荣[1] 舒忠[1] TIAN Xue-jun;ZHANG Wen-qiang;MA Zi-xuan;CHEN Liang-zhe;YE Hui-rong;SHU Zhong(College of New Energy,Jingchu University of Technology,Jingmen 448000,China)

机构地区：[1]荆楚理工学院新能源学院,荆门448000

出　　处：《印刷与数字媒体技术研究》2024年第5期33-44,共12页Printing and Digital Media Technology Study

基　　金：湖北省高等学校优秀中青年科技创新团队计划项目(No.T2022038);荆楚理工学院校级科研团队项目(No.TD202101);2023年度荆门市科技计划重点项目(No.2023YFZD056);荆门市课程思政研究项目(No.JMG2024141)。

摘　　要：为了更好地进行图像重建,加强对图像多尺度特征表示和特征融合的处理,本研究提出一种新的多尺度特征融合图像重建网络模型。模型包括迭代降采样和迭代上采样过程。迭代降采样过程是通过在拉普拉斯金字塔残差控制网络模型中将高斯卷积核与子采样和高斯平滑滤波迭代操作规则相结合完成的。迭代上采样过程是通过使用拉普拉斯卷积核和二阶差分操作规则实现的。GJ-UNet深度学习网络模型通过其编码器下采样模块实现图像多尺度语义特征的精细分类,并在解码器上采样模块中应用反卷积和卷积操作规则,规范处理图像多尺度语义特征。实验表明,所提出的方法可以实现高精度的特征提取,同时对于图像特征融合的相关性更强,提取的图像边缘信息更清晰且相对噪声信息更低,重建图像的视觉效果基本与原始输入图像相同。本研究有望广泛应用于计算机图像视觉领域。Aimed to better perform image reconstruction and enhance the processing of multi-scale feature representation and fusion of images,a new multi-scale feature fusion image reconstruction network model was proposed in this study.The model included iterative downsampling and iterative upsampling processes.The iterative downsampling process was completed by combining the Gaussian convolution kernel with the iterative operation rules of subsampling and Gaussian smooth filtering in the Laplacian pyramid residual control network model.The iterative upsampling process was achieved by using Laplace convolution kernels and second-order differential operation rules.The GJ-UNet deep learning network model achieved fine classification of image multi-scale semantic features through its encoder downsampling module,and applied deconvolution and convolution operation rules in the decoder upsampling module to standardize the processing of image multi-scale semantic features.The results showed that the proposed method can achieve high-precision feature extraction,and has stronger correlation with image feature fusion.The extracted image edge information is clearer and has lower relative noise information.The visual effect of the reconstructed image is basically the same as the original input image.This study is expected to be widely used in the field of computer image vision.

关 键 词：压缩感知图像重构 图像多尺度特征 拉普拉斯金字塔模型 差分运算 GJ-UNet深度学习网络模型 Dice损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术] TS80[轻工技术与工程]