作者：许富景¹, 陈长颖², 杜少成¹

作者单位：1. 山西大学自动化与软件学院，山西 太原 030013;
2. 山西大学数学科学学院，山西 太原 030006

关键词：压缩感知;图像重建;卷积神经网络;子像素卷积;残差网络

摘要：

传统压缩感知图像重建方法存在着运算量大、重建图像过程耗时以及低采样率下重建图像精度低等问题。针对这些问题，提出一种基于改进CNN的压缩感知自然图像重建方法，该方法主要包括：线性映射网络、初始重建网络和最终重建网络。首先，利用线性映射卷积网络自适应获得测量向量；其次，把通过学习得来的自适应测量向量输入到初始重建子像素卷积网络中，利用子像素卷积网络对上一步得到的特征图进行低分辨率初始重建；最后，利用残差网络对重建图像进行高精度重建，从而有效提升重建图像的质量。大量实验结果表明，该方法在PSNR、SSIM和视觉效果这三方面都明显优于其他方法，在采样率为0.0625时它的PSNR和SSIM的平均结果分别为31.55 dB和0.9253。

Compressed sensing natural image reconstruction method based on improved CNN
XU Fujing¹, CHEN Changying², DU Shaocheng¹
1. School of Automation and Software Engineering, Shanxi University, Taiyuan 030013, China;
2. School of Mathematical Sciences, Shanxi University, Taiyuan 030006, China
Abstract: Traditional compressed sensing image reconstruction methods have many problems, such as large amount of computation, time-consuming reconstruction process and low accuracy reconstruction under low sampling rate. To solve these problems, a compressed sensing reconstruction method of natural image based on improved CNN is proposed, which mainly includes: Linear mapping network, initial and final reconstruction network reconstruction it first using linear mapping convolution network adaptive gain measurement vector, secondly, through the study of adaptive measurement vector input to the original reconstruction convolution sub pixel in the network, using the sub pixel characteristic figure obtained by convolution network to step on low resolution initial reconstruction, in the end, The residual network is used to reconstruct the reconstructed image with high precision so as to improve the quality of the reconstructed image effectively. A large number of experimental results show that this method is significantly superior to other methods in PSNR, SSIM and visual effect. When the sampling rate is 0.0625, the average results of PSNR and SSIM are 31.55 dB and 0.9253, respectively.

Keywords: compressed sensing;image reconstruction;convolutional neural network;sub-pixel convolution;residual network
2022, 48(9):7-16  收稿日期: 2021-07-18;收到修改稿日期: 2021-10-13
基金项目: 国家自然科学基金（61903240）
作者简介: 许富景（1989-），男，山西大同市人，副教授，硕士生导师，博士，研究方向为动态测控与智能仪器、智能物联网技术
参考文献
[1] MUN S, FOWLER J E. DPCM for quantized block-based compressed sensing of images[C]//Signal Processing Conference. IEEE, 2012: 1424-1428.
[2] DU J, XIE X, WANG C, et al. Fully convolutional measurement network for compressive sensing image reconstruction[J]. Neurocomputing, 2019, 328: 105-112
[3] 张从华, 刘景鑫, 杨勇, 等. 医用诊断X射线数字图像评价模体与轮廓提取研究[J]. 中国测试, 2018, 44(8): 97-101
[4] MITRA U, CHOUDHARY S, HOVER F, et al. Structured sparse methods for active ocean observation systems with communication constraints[J]. IEEE Communications Magazine, 2015, 53(11): 88-96
[5] WANG L, LU K, LIU P. Compressed sensing of a remote sensing image based on the priors of the reference image[J]. IEEE Geoscience & Remote Sensing Letters, 2014, 12(4): 736-740
[6] 董成举, 刘文威, 李小兵, 等. 六轴工业机器人工作空间分析及区域性能研究[J]. 中国测试, 2020, 46(5): 154-160
[7] SANKARANARAYANAN A C, STUDER C, BARANIUK R G. CS-MUVI: Video compressive sensing for spatial-multiplexing cameras[C]//2012 IEEE International Conference on Computational Photography (ICCP). IEEE, 2012: 1-10.
[8] LI S, XU L D, WANG X. Compressed sensing signal and data acquisition in wireless sensor networks and internet of things[J]. IEEE Transactions on Industrial Informatics, 2013, 9(4): 2177-2186
[9] GAN L. Block compressed sensing of natural images[C]//International Conference on Digital Signal Processing. IEEE, 2007: 403–406.
[10] DINH K Q, SHIM H J, JEON B. Measurement coding for compressive imaging using a structural measuremnet matrix[C]//2013 IEEE International Conference on Image Processing. IEEE, 2014: 10-13.
[11] GAO X, JIAN Z, CHE W, et al. Block-based compressive sensing coding of natural images by local structural measurement matrix[C]//2015 Data Compression Conference. IEEE, 2015: 133-142.
[12] MOUSAVI A, PATEL A B, BARANIUK R G. A deep learning approach to structured signal recovery[C]//2015 53rd Annual Allerton Conference on Communication, Control, and Computing, 2016: 1336-1343.
[13] KULDEEP K, SUHAS L, PAVAN T, et al. ReconNet: non-iterative reconstruction of images from compressively sensed measurements[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2016: 449-458.
[14] YAO H, DAI F, ZHANG S, et al. DR2-Net: deep residual reconstruction network for image compressive sensing[J]. Neurocomputing, 2019, 359: 483-493
[15] CHAO D, CHEN C L, KAIMING H, et al. Image super-resolution using deep convolutional networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38(2): 295-307
[16] WANG Z, DING L, CHANG S, et al. D3: deep dual-domain based fast restoration of jpeg-compressed images[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016: 2764-2772.
[17] XIE X, WANG Y, SHI G, et al. Adaptive measurement network for CS image reconstruction[J]. Computer Vision, 2017, 772: 407-417
[18] SHEN Y, LI J, ZHU Z, et al. Image reconstruction algorithm from compressed sensing measurements by dictionary learning[J]. Neurocomputing, 2015, 151: 1153-1162
[19] BARANIUK R G, CEVHER V, DUARTE M F, et al. Model-based compressive sensing[J]. IEEE Transactions on Information Theory, 2010, 56(4): 1982-2001
[20] TROPP J A and GILBERT A C. Signal recovery from random measurements via Orthogonal Matching Pursuit[J]. IEEE Transactions on Information Theory, 2007, 53(12): 4655-4666
[21] FIGUEIREDO M, NOWAK R D, WRIGHT S J. Gradient projection for sparse reconstruction: application to compressed sensing and other inverse problems[J]. IEEE Journal of Selected Topics in Signal Processing, 2008, 1(4): 586-597
[22] ZHU T. New over-relaxed monotone fast iterative shrinkage-thresholding algorithm for linear inverse problems[J]. IET Image Processing, 2019, 13(14): 2888-2896
[23] ZHOU Y, ZHANG M, ZHU J, et al. A randomized block-coordinate adam online learning optimization algorithm[J]. Neural Computing and Applications, 2020, 32(16): 12671-12684
[24] SHI W, CABALLERO J, HUSZAR F, et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016: 1874–1883.
[25] CONG B, LING H, XIANG P, et al. Optimization of deep convolutional neural network for large scale image retrieval[J]. Neurocomputing, 2018, 303: 60-67
[26] IOFFE S, SZEGEDY C. Batch Normalization: accelerating deep network training by reducing internal covariate shift[C]//International Conference on Machine Learning, 2015: 448-456.
[27] BEVILACQUA M, ROUMY A, GUILEMOT C, et al. Low-complexity single image super-resolution based on nonnegative neighbor embedding[C]//2012 23rd British Machine Vision Conference. 2012: 135.1-135.10.
[28] ZEYDE R, ELAD M, PROTTER M. On single image scale-up using sparse-representations[C]//Int. Conf. Curves Surfaces. 2010: 711–730.