中国物理B ›› 2020, Vol. 29 ›› Issue (5): 54203-054203.doi: 10.1088/1674-1056/ab7b4e

• ELECTROMAGNETISM, OPTICS, ACOUSTICS, HEAT TRANSFER, CLASSICAL MECHANICS, AND FLUID DYNAMICS • 上一篇    下一篇

An image compressed sensing algorithm based on adaptive nonlinear network

Yuan Guo(郭媛), Wei Chen(陈炜), Shi-Wei Jing(敬世伟)   

  1. School of Computer and Control Engineering, Qiqihar University, Qiqihar 161006, China
  • 收稿日期:2019-11-01 修回日期:2020-02-01 出版日期:2020-05-05 发布日期:2020-05-05
  • 通讯作者: Wei Chen E-mail:1010172469@qq.com
  • 基金资助:
    Project supported by the National Natural Science Foundation of China (Grant No. 61872204), the Natural Science Fund of Heilongjiang Province, China (Grant No. F2017029), the Scientific Research Project of Heilongjiang Provincial Universities, China (Grant No. 135109236), and the Graduate Research Project, China (Grant No. YJSCX2019042).

An image compressed sensing algorithm based on adaptive nonlinear network

Yuan Guo(郭媛), Wei Chen(陈炜), Shi-Wei Jing(敬世伟)   

  1. School of Computer and Control Engineering, Qiqihar University, Qiqihar 161006, China
  • Received:2019-11-01 Revised:2020-02-01 Online:2020-05-05 Published:2020-05-05
  • Contact: Wei Chen E-mail:1010172469@qq.com
  • Supported by:
    Project supported by the National Natural Science Foundation of China (Grant No. 61872204), the Natural Science Fund of Heilongjiang Province, China (Grant No. F2017029), the Scientific Research Project of Heilongjiang Provincial Universities, China (Grant No. 135109236), and the Graduate Research Project, China (Grant No. YJSCX2019042).

摘要: Traditional compressed sensing algorithm is used to reconstruct images by iteratively optimizing a small number of measured values. The computation is complex and the reconstruction time is long. The deep learning-based compressed sensing algorithm can greatly shorten the reconstruction time, but the algorithm emphasis is placed on reconstructing the network part mostly. The random measurement matrix cannot measure the image features well, which leads the reconstructed image quality to be improved limitedly. Two kinds of networks are proposed for solving this problem. The first one is ReconNet's improved network IReconNet, which replaces the traditional linear random measurement matrix with an adaptive nonlinear measurement network. The reconstruction quality and anti-noise performance are greatly improved. Because the measured values extracted by the measurement network also retain the characteristics of image spatial information, the image is reconstructed by bilinear interpolation algorithm (Bilinear) and dilate convolution. Therefore a second network USDCNN is proposed. On the BSD500 dataset, the sampling rates are 0.25, 0.10, 0.04, and 0.01, the average peak signal-noise ratio (PSNR) of USDCNN is 1.62 dB, 1.31 dB, 1.47 dB, and 1.95 dB higher than that of MSRNet. Experiments show the average reconstruction time of USDCNN is 0.2705 s, 0.3671 s, 0.3602 s, and 0.3929 s faster than that of ReconNet. Moreover, there is also a great advantage in anti-noise performance.

关键词: compressed sensing, deep learning, bilinear interpolation, dilate convolution

Abstract: Traditional compressed sensing algorithm is used to reconstruct images by iteratively optimizing a small number of measured values. The computation is complex and the reconstruction time is long. The deep learning-based compressed sensing algorithm can greatly shorten the reconstruction time, but the algorithm emphasis is placed on reconstructing the network part mostly. The random measurement matrix cannot measure the image features well, which leads the reconstructed image quality to be improved limitedly. Two kinds of networks are proposed for solving this problem. The first one is ReconNet's improved network IReconNet, which replaces the traditional linear random measurement matrix with an adaptive nonlinear measurement network. The reconstruction quality and anti-noise performance are greatly improved. Because the measured values extracted by the measurement network also retain the characteristics of image spatial information, the image is reconstructed by bilinear interpolation algorithm (Bilinear) and dilate convolution. Therefore a second network USDCNN is proposed. On the BSD500 dataset, the sampling rates are 0.25, 0.10, 0.04, and 0.01, the average peak signal-noise ratio (PSNR) of USDCNN is 1.62 dB, 1.31 dB, 1.47 dB, and 1.95 dB higher than that of MSRNet. Experiments show the average reconstruction time of USDCNN is 0.2705 s, 0.3671 s, 0.3602 s, and 0.3929 s faster than that of ReconNet. Moreover, there is also a great advantage in anti-noise performance.

Key words: compressed sensing, deep learning, bilinear interpolation, dilate convolution

中图分类号:  (Image reconstruction; tomography)

  • 42.30.Wb
42.68.Sq (Image transmission and formation)