摘要
针对图像在细节丰富和结构复杂区域重建效果不佳的问题,提出一种对称融合多尺度双域网络(SFMD-Net),通过融合频域和空间域信息,提高图像重建质量。在频域增强支路,利用二维离散余弦变换获取频域信息,通过双残差注意模块深化频域关键特征的学习。在空域多尺度重建支路,设计多尺度深度细化模块,加强对空间域内不同层次和尺度特征的提取与增强。引入对称融合策略确保频域与空间域特征的有效融合,提高重建图像的细节表现和整体视觉质量。实验结果表明,与当前主流压缩感知重建算法相比,所提算法在性能和效果上均有改进,尤其是在重建边缘明显和纹理丰富的图像方面。
Aiming at the problem of poor image reconstruction in detail-rich and structurally complex regions,a symmetric fusion multi-scale dual-domain network(SFMD-Net)was proposed to improve the quality of image reconstruction by innovatively merging frequency and spatial domain information.In the frequency domain enhancement branch,the two-dimensional discrete cosine transform was utilized to acquire frequency domain information,and a dual residual attention module was employed to deepen the learning of key features in the frequency domain.In the spatial domain multi-scale reconstruction branch,a multi-scale deep refinement module was designed to enhance the extraction and enhancement of features at different levels and scales within the spatial domain.A symmetric fusion strategy was introduced to ensure the effective fusion of frequency and spatial domain features,which improved the detailed performance and overall visual quality of the reconstructed images.Experimental validation demonstrates that the proposed algorithm exhibits improvements in performance and effectiveness compared to current mainstream compressive sensing reconstruction algorithms,particularly in reconstructing images with pronounced edges and rich textures.
作者
林乐平
廖心竹
欧阳宁
LIN Le-ping;LIAO Xin-zhu;OUYANG Ning(School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China)
出处
《计算机工程与设计》
北大核心
2025年第7期1905-1911,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(62001133)
广西科技基地和人才专项基金项目(桂科AD19110060)
广西高等学校千名中青年骨干教师培育计划基金项目
广西自然科学基金项目(2017GXNSFBA198212)
广西无线宽带通信与信号处理重点实验室基金项目(GXKL06200114)。
关键词
压缩感知
多尺度特征
离散余弦变换
图像重建
对称融合
深度学习
双域网络
compressed sensing
multi-scale features
discrete cosine transform
image reconstruction
symmetric fusion
deep learning
dual-domain network
作者简介
林乐平(1980-),女,广西桂平人,博士,教授,研究方向为机器学习、智能信息处理、图像信号处理;廖心竹(2000-),女,广西桂林人,硕士研究生,研究方向为图像信号处理;通讯作者:欧阳宁(1972-),男,湖南永州人,硕士,教授,研究方向为数字图像处理、智能信息处理。E-mail:ouyangning@guet.edu.cn。
