School of Information Engineering, East China Jiaotong University, Nanchang 330013, China
Show less
文章历史+
Received
Accepted
Published
2023-07-09
Issue Date
2025-10-27
PDF (4734K)
摘要
图像压缩感知能从低采样观测中重建出高质量图像。将深度学习应用于图像压缩感知,可显著提高图像重建质量。然而,基于深度学习的图像压缩感知方法存在模型可解释性差、结构盲目设计而影响重建性能的问题。针对这些问题,提出了一种基于零值域分解的深度图像压缩感知方法(range‒null space decomposition based deep image compressive sensing network,RND‒Net)。该方法通过全局卷积采样的方式稀疏感知图像的特征信息,通过学习信号相关的采样矩阵,使采样值包含更丰富的图像特征,且相较一般的逐块采样方式,在全局层面上的采样可明显减少块状伪影;基于零值域分解的数学表示,将采样与重建过程转化为端到端深度学习模型,借助深度神经网络拟合所涉及的线性或非线性运算,相比传统方法缩短了模型推理时间,提升了图像重建能力。上述将数学先验知识有效融入数据驱动的方法称为协同驱动,既充分利用了数学先验知识,强化了模型的可解释性,使模型结构更易于设计,又发挥了以深度学习为代表的数据驱动方法的自主寻优能力,相比其他深度压缩感知方法更易于获得全局最优解。在多个测试集上的实验证明,RND‒Net与目前图像重建能力较好的算法相比显著提升了图像重建质量,减少了单幅图像重建时间。当采样率为0.1、测试集为BSDS68时,RND‒Net比AutoBCS在峰值信噪比(PSNR)上平均高1.02 dB。在测试集Set14上,RND‒Net对于混合驱动的GPX‒ADMM‒Net的平均PSNR和结构相似性指数(SSIM)增益分别为1.15 dB和0.051 8;重建单幅图像时,RND‒Net比GPX‒ADMM‒Net快约0.104 9 s。
Abstract
Image compressive sensing (ICS) reconstructs high-quality images from low-sampling observations. Applying deep learning to ICS significantly improves image reconstruction quality. However, deep learning-based ICS methods suffer from poor model interpretability and blind structural design, which degrade reconstruction performance. Therefore, a Range-Null Space Decomposition(RND)-based Deep Image Compressive Sensing Network is proposed, referred to as RND‒Net. This method sparsely senses image feature information through global convolutional sampling. Learning signal-related sampling matrices enables the sampling values to contain richer image features. Compared to common block-wise sampling methods, the global-level sampling approach remarkably reduces block artifacts. In addition, based on the mathematical representation of RND, the sampling and reconstruction processes are transformed into an end-to-end deep learning model. With the deep neural network fitting linear or nonlinear operations involved, model reasoning time is reduced, and image reconstruction ability is improved compared to traditional algorithms. This method, which effectively integrates mathematical prior knowledge into data-driven approaches, is called a collaborative driving method. It fully utilizes mathematical prior knowledge, strengthens model interpretability, simplifies model structure design, and uses the autonomous optimization capabilities of data-driven methods represented by deep learning. Compared to other deep compressive sensing methods, it more easily reaches the global optimal solution. Experiments on multiple test sets demonstrate that RND‒Net significantly improves image reconstruction quality and reduces the time required to reconstruct a single image compared to state-of-the-art algorithms. When the sampling rate is 0.1, and the test set is BSDS68, the average PSNR of RND‒Net is about 1.02 dB higher than that of AutoBCS. On Set14, the average PSNR and SSIM gains of RND‒Net over hybrid-driven GPX‒ADMM‒Net are 1.15 dB and 0.051 8, respectively. In addition, RND‒Net is about 0.104 9 s faster than GPX‒ADMM‒Net when reconstructing a single image.
图像压缩感知(ICS)是压缩感知体系中的重要研究问题。ICS的一般模型需明确采样矩阵、稀疏变换及非线性重建方法等要素。传统上,采样矩阵应使被采样图像得到充分采样,以便之后的重建,一般采用高斯随机矩阵(GRM)、随机伯努利矩阵等,然而,这些采样矩阵都是信号无关的,由其得到的采样值缺乏能够高质量恢复原始图像的充分的特征信息[3]。稀疏变换的作用是使输入图像 x 的像素矩阵稀疏化,且越彻底的稀疏越有助于图像的恢复重建[6],传统的有多假设法(MH)[7]、全变分法(TVAL3)[8]等。此后,Zhang等[9]提出了组稀疏(GSR)方法,在图像重建精度上取得了一定提升,但上述方法的重建速度慢是不容忽视的问题。在重建层面,经典方法有匹配追踪(MP)[10]、迭代收缩阈值算法(ISTA)[11]等。这些方法能够充分利用图像先验知识,在像素空间对每一像素点都进行比较精确地重建,不过,存在由反复迭代引起的计算开销大、单幅图像重建速度慢等问题。
本文将深度学习方法和数学先验知识或优化理论融合为一体,既凭借神经网络模型优越的特征提取和重建性能增强了后者的数据拟合、变量寻优的能力,使其模型求解过程高效进行,又弥补了一般深度学习方法的可解释性不足,更克服了多数纯深度学习方法未充分利用图像先验知识、已有混合驱动算法仅对旧有网络模块进行简单机械地组合的缺点,本文称其为模型和数据协同驱动方法。根据协同驱动的思路,在零值域分解(RND)[20‒21]的启发下,将深度学习方法与零值域分解结合,集成数学先验知识和数据驱动方法在ICS方向的优势,构建基于零值域分解的深度图像压缩感知模型(range‒null space decomposition based deep image compressive sensing network,RND‒Net)。该模型用全局卷积采样的方式学习与信号相关的采样矩阵,相较许多工作使用的逐块卷积采样方式可增大模型感受野,从而获取包含丰富信息的采样值;之后,通过端到端的方式将零值域分解的步骤网络化,设计了级联残差块及多级长短跳跃连接的网络模型以得到零域提取项,且将采样信息经捕捉某一尺度特征后加以多次复用,有助于发挥采样值所蕴含信息的更大作用,并通过线性卷积运算获取图像的低频与高频信息(即值域和零域),将二者融合后得到最终重建图像。
式中:,,为图像块索引,;;且,为采样率,。分析上述矩阵相乘过程可以发现,采样后得到的输出值的每一个元素值由 A 的对应行向量与逐元素相乘再求和后得到,若将 A 的每一行视作一个滤波器并用CNN模拟,则构造M个滤波器(卷积核)即可完成对图像块的采样[3]。经典的CSNet[3]的采样模块即使用上述逐块采样方式。不过,逐块采样所使用的卷积核尺寸过大且仅经过单层卷积层处理,使采样模块的计算复杂度较大,同时不利于获取高级特征信息。若将单卷积层改为多卷积网络,适度增加网络深度,既能完成采样过程,又可在增大模型感受野的同时获得原图像的高级特征,有利于对特征信息进行重建。事实上,逐块卷积采样所用的大尺寸卷积核可替换为多个小尺寸卷积核级联,实现与其等同的效果。
ShannonC E.Communication in the presence of noise[J].Proceedings of the IRE,1949,37(1):10‒21. doi:10.1109/jrproc.1949.232969
[2]
YangYan, SunJian, LiHuibin,et al.ADMM-CSNet:A deep learning approach for image compressive sensing[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2020,42(3):521‒538. doi:10.1109/tpami.2018.2883941
[3]
ShiWuzhen, JiangFeng, LiuShaohui,et al.Image compr-essed sensing using convolutional neural network[J].IEEE Transactions on Image Processing,2019,29:375‒388. doi:10.1109/tip.2019.2928136
[4]
DonohoD L.Compressed sensing[J].IEEE Transactions on Information Theory,2006,52(4):1289‒1306. doi:10.1109/tit.2006.871582
[5]
WenBihan, RavishankarS, PfisterL,et al.Transform learning for magnetic resonance image reconstruction:From mo-del-based learning to building neural networks[J].IEEE Si-gnal Processing Magazine,2020,37(1):41‒53. doi:10.1109/msp.2019.2951469
[6]
ZhuLu, SongChao, LiuYuanyuan,et al.Microwave radiation image reconstruction method based on the mixed sp-arse basis dictionary learning[J].Journal of Electronics & Information Technology,2016,38(11):2724‒2730.
ChenChen, TramelE W, FowlerJ E.Compressed-sensing re-covery of images and video using multihypothesis predictions[C]//Proceedings of the 2011 Conference Record of the Forty Fifth Asilomar Conference on Signals,Systems and Computers(ASILOMAR).Pacific Grove:IEEE,2011:1193‒1198. doi:10.1109/acssc.2011.6190204
[9]
WangZhangyang, LiuDing, ChangShiyu,et al .D 3:Deep dual-domain based fast restoration of JPEG-compressed images[C]//Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition(CVPR).Las Vegas:IEEE,2016:2764‒2772. doi:10.1109/cvpr.2016.302
[10]
ZhangJian, ZhaoDebin, GaoWen.Group-based sparse representation for image restoration[J].IEEE Transactions on Image Processing,2014,23(8):3336‒3351. doi:10.1109/tip.2014.2323127
[11]
MallatS G, ZhangZhifeng.Matching pursuits with time-frequency dictionaries[J].IEEE Transactions on Signal Pr-ocessing,1993,41(12):3397‒3415. doi:10.1109/78.258082
[12]
BeckA, TeboulleM.A fast iterative shrinkage-thresholding algorithm for linear inverse problems[J].SIAM Journal on Imaging Sciences,2009,2(1):183‒202. doi:10.1137/080716542
[13]
YuRui, LiuYuanyuan, ZhuLu.Inverse design of high deg-ree of freedom meta-atoms based on machine learning and genetic algorithm methods[J].Optics Express,2022,30(20):35776‒35791. doi:10.1364/oe.472280
[14]
MousaviA, PatelA B, BaraniukR G.A deep learning approach to structured signal recovery[C]//Proceedings of the 2015 53rd Annual Allerton Conference on Communication,Control,and Computing(Allerton).Monticello:IEEE,2015:1336‒1343. doi:10.1109/allerton.2015.7447163
[15]
KulkarniK, LohitS, TuragaP,et al.ReconNet:Non-iterative reconstruction of images from compressively sensed measurements[C]//Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition(CVPR).Las Vegas:IEEE,2016:449‒458. doi:10.1109/cvpr.2016.55
[16]
FanZien, LianFeng, QuanJiani.Global sensing and measurements reuse for image compressed sensing[C]//Proce-edings of the 2022 IEEE/CVF Conference on Computer Vi-sion and Pattern Recognition(CVPR).New Orleans:IEEE,2022:8944‒8953. doi:10.1109/cvpr52688.2022.00875
[17]
ZengChunyan, YanKang, WangZhifeng,et al.Multi-scale generative adversarial network for image compressed sen-sing and reconstruction[J].Journal of Jilin University(Engineering and Technology Edition),2023,53(10):2923‒2931.
YangYan, SunJian, LiHhuibin,et al.Deep ADMM-Net for compressive sensing MRI[J].Advances in Neural Information Processing Systems,2016,29:1‒9.
[20]
ZhongYuanhong, ZhouYujie, ZhangJing,et al.Deep compressive sensing image reconstruction network based on non-local prior[J].Journal of Electronics & Information Technology,2023,45(2):654‒663.
WangYinhuai, HuYujie, YuJiwen,et al.GAN prior based null-space learning for consistent super-resolution[EB/OL].(2022‒11‒24)[2023‒03‒23].doi:10.1609/aaai.v37i3.25372
[25]
WangYinhuai, YuJiwen, ZhangJian.Zero-shot image restoration using denoising diffusion null-space model[EB/OL].(2022‒12‒01)[2023‒03‒23].
[26]
CandesE J, WakinM B.An introduction to compressive sampling[J].IEEE Signal Processing Magazine,2008,25(2):21‒30. doi:10.1109/msp.2007.914731
[27]
CandesE J, TaoT.Decoding by linear programming[J].IEEE Transactions on Information Theory,2005,51(12):4203‒4215. doi:10.1109/tit.2005.858979
[28]
XieYahong, WangHailin, WangJianjun.CMCS-net:Image compressed sensing with convolutional measurement via DCNN[J].IET Image Processing,2020,14(15):3839‒3850. doi:10.1049/iet-ipr.2020.0834
[29]
ChenZan, GuoWenlong, FengYuanjing,et al.Deep-learned regularization and proximal operator for image compressive sensing[J].IEEE Transactions on Image Processing,2021,30:7112‒7126. doi:10.1109/tip.2021.3088611
[30]
ZhangJian, ZhaoChen, GaoWen.Optimization-inspired co-mpact deep compressive sensing[J].IEEE Journal of Selec-ted Topics in Signal Processing,2020,14(4):765‒774. doi:10.1109/jstsp.2020.2977507
[31]
ArbeláezP, MaireM, FowlkesC,et al.Contour detection and hierarchical image segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2011,33(5):898‒916. doi:10.1109/tpami.2010.161
[32]
BevilacquaM, RoumyA, GuillemotC,et al.Low-complexity single-image super-resolution based on nonnegative neighbor embedding[C]//Proceedings of the British Machine Vision Conference 2012.Surrey:British Machine Vision Association,2012:1‒10. doi:10.5244/c.26.135
[33]
ZeydeR, EladM, ProtterM.On single image scale-up using sparse-representations[C]//Curves and Surfaces.Berlin:Springer Berlin Heidelberg,2012:711‒730. doi:10.1007/978-3-642-27413-8_47
[34]
PengYang, TanHanlin, LiuYu,et al.Structure prior guided deep network for compressive sensing image reconstruction from big data[C]//Proceedings of the 2020 6th International Conference on Big Data and Information Analytics (BigDIA).Shenzhen:IEEE,2020:270‒277. doi:10.1109/bigdia51454.2020.00050
[35]
ZhangJian, GhanemB.ISTA-net:Interpretable optimization-inspired deep network for image compressive sensing[C]//Proceedings of the 2018 IEEE/CVF Conference on Comp-uter Vision and Pattern Recognition.Salt Lake City:IEEE,2018:1828‒1837. doi:10.1109/cvpr.2018.00196
[36]
CuiWenxue, LiuShaohui, JiangFeng,et al.Image compres-sed sensing using non-local neural network[J].IEEE Transactions on Multimedia,2021,25:816‒830.