摘要
基于卷积神经网络(Convolutional Neural Networks,CNN)的图像压缩感知重构算法难以捕捉高分辨率图像的长距离依赖关系,采用Transformer虽能解决该问题,但网络参数量和图像重构时间成倍增长。基于此,本文提出了一种联合图像层级特征的压缩感知迭代重构网络(Combining Image Hierarchical-Feature Network,CHFNet),在提高图像重构质量的同时减少重构时间。CHFNet由采样和重构两个子网络组成,采样子网络通过可学习的采样矩阵为重构过程提供更有效的测量值。在重构子网络中,设计了一种使用梯度下降操作和特征优化操作的迭代策略,同时提出一种轻量级CNN-Transformer混合架构,能够建模并优化高细粒度的图像层级特征,在增强网络感知能力的同时降低计算复杂度。此外,CHFNet通过联合优化学习采样重构,实现了完整的端到端训练。实验结果表明,所提算法在多个公共基准数据集上取得了良好的重构效果。在Urban100数据集上,相较于现有最优算法CSformer,平均PSNR,SSIM分别提升0.63 dB和0.0076;在0.10采样率下,相较CSformer在Set11,BSD68和Urban100数据集上的平均重构时间分别减少了2.7447 s,3.5510 s和4.7750 s。
The compressive sensing image reconstruction algorithms based on Convolutional Neural Net⁃works could not capture long-range dependency of high-resolution images.Although Transformer can ad⁃dress this issue,it significantly increases the number of network parameters and the image reconstruction time.This paper proposed CHFNet,a combining image hierarchical-feature network for compressive sensing iterative-reconstruction to improve image reconstruction quality and reduce reconstruction time.CHFNet consisted of two sub-networks,sampling and reconstruction.The sampling sub-network utilized a learnable sampling matrix to provide more effective measurements for reconstruction phase.In the recon⁃struction sub-network,we introduced an iterative strategy using gradient descent and feature optimization operations,and proposed a lightweight CNN-Transformer hybrid architecture to model and optimize extremely fine-grained image hierarchical-feature,enhancing network’s sensing-capability and reducing com⁃putation complexity.Moreover,CHFNet achieved complete end-to-end training by jointly optimizing sam⁃pling-reconstruction process.The experimental results show that the proposed algorithm obtains satisfacto⁃ry recovery performance on several public benchmark datasets.On the Urban100 dataset,the method of this paper improves the average PSNR and SSIM metrics by 0.63 dB and 0.0076 respectively compared to the existing optimal algorithm CSformer.At 0.10 sampling rate,the average reconstruction time of CHFNet decreases 2.7447 s,3.5510 s,and 4.7750 s compared to CSformer on Set11,BSD68,and Urban100 datasets respectively.
作者
刘玉红
杨恒
LIU Yuhong;YANG Heng(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2024年第14期2311-2324,共14页
Optics and Precision Engineering
基金
国家自然科学基金资助项目(No.62161016,No.61661025)。
