• Optics and Precision Engineering
  • Vol. 31, Issue 14, 2135 (2023)
Yuanhong ZHONG1,*, Qianfeng XU1, Yujie ZHOU1, and Shanshan WANG2
Author Affiliations
  • 1School of Microelectronics and Communication Engineering, Chongqing University, Chongqing400044, China
  • 2Institute of Physical Science and Information Technology, Anhui University, Hefei30039, China
  • show less
    DOI: 10.37188/OPE.20233114.2135 Cite this Article
    Yuanhong ZHONG, Qianfeng XU, Yujie ZHOU, Shanshan WANG. Image reconstruction based on deep compressive sensing combined with global and local features[J]. Optics and Precision Engineering, 2023, 31(14): 2135 Copy Citation Text show less
    Global-to-Local Compressive Sensing Image Reconstruction Model Structure
    Fig. 1. Global-to-Local Compressive Sensing Image Reconstruction Model Structure
    8 test images from Set11[32]
    Fig. 2. 8 test images from Set1132
    Sampling rate is 10%, and the reconstruction images of each algorithm on the image House are compared
    Fig. 3. Sampling rate is 10%, and the reconstruction images of each algorithm on the image House are compared
    Sampling rate is 20%, and the reconstruction image comparison of each algorithm on the image Monarch
    Fig. 4. Sampling rate is 20%, and the reconstruction image comparison of each algorithm on the image Monarch
    Change curve of loss with the number of training iterations (epochs) at 20% sampling rate
    Fig. 5. Change curve of loss with the number of training iterations (epochs) at 20% sampling rate
    G2LNet reconstruction image and filter flow visualization at 30% sampling rate
    Fig. 6. G2LNet reconstruction image and filter flow visualization at 30% sampling rate
    采样率算法MonarchHouseBarbaraLenaParrotsPeppersBoatsC.man
    10%MH2223.1930.3026.7326.1225.3526.0026.1122.12
    ReconNet2321.5126.6922.5024.4723.2322.6724.1521.66
    CSNet926.7331.6824.2428.5727.4026.6628.8024.92
    ISTA-Net+[2425.7230.4923.5227.5026.3727.1327.4123.76
    G2LNet28.0732.0025.2328.8628.8827.6128.9525.46
    20%MH2227.1033.8430.8129.8129.2129.7829.9125.87
    ReconNet2322.8927.9522.8725.3924.5624.0425.9822.64
    CSNet929.5733.4224.9830.7529.7728.4230.9726.79
    ISTA-Net+[2431.0134.9926.7831.1429.9630.4531.9127.65
    G2LNet31.6135.1627.2731.7131.1231.3832.4428.30
    30%MH2229.2035.6932.9931.9931.0031.3532.2528.08
    ReconNet2329.2133.6125.6533.7426.8829.7730.2026.90
    CSNet932.5836.4728.3833.3532.8730.7533.2128.64
    ISTA-Net+[2434.8037.0730.1333.4532.9135.5435.2230.35
    G2LNet35.2737.5630.8734.2633.5835.8135.8731.28
    Table 1. PSNR of reconstructed images of different algorithms in 8 pictures
    采样率数据集MH22ReconNet23CSNet9ISTA-Net+[24G2LNet
    10%Set528.5624.3131.5428.6130.56
    Set1125.8222.4527.3726.4927.95
    BSD6823.9723.6226.1225.3026.66
    Avg.26.1223.4628.3426.8028.39
    20%Set526.1223.4628.3426.8028.39
    Set1128.9324.4429.3330.8031.21
    BSD6826.9825.1229.5129.0329.59
    Avg.27.3424.3429.0628.8829.73
    30%Set534.0927.7836.6935.4536.03
    Set1131.5525.7430.9833.7033.71
    BSD6828.7926.2731.0130.3531.29
    Avg.31.4826.6032.8933.1733.68
    Table 2. PSNR of reconstructed images of different algorithms in multiple datasets
    采样率数据集MH22ReconNet23CSNet9ISTA-Net+[24G2LNet
    10%Set50.835 00.734 10.901 00.839 80.898 6
    Set110.782 20.624 50.841 20.803 60.868 7
    BSD680.655 90.648 20.772 40.700 10.789 6
    Avg.0.757 70.668 90.838 20.781 20.852 3
    20%Set50.888 10.794 00.950 10.902 10.943 6
    Set110.872 10.732 40.901 20.913 10.922 5
    BSD680.769 30.705 20.875 70.864 20.882 0
    Avg.0.843 20.743 90.909 00.893 10.916 0
    30%Set50.918 50.823 30.963 30.934 10.959 3
    Set110.906 30.804 60.891 20.938 20.950 0
    BSD680.827 60.784 20.882 40.878 20.918 6
    Avg.0.884 10.804 00.912 30.916 80.942 6
    Table 3. Comparison results of SSIM of reconstructed images with different algorithms
    数据集算法模型采样率
    10%20%30%
    PSNRSSIMPSNRSSIMPSNRSSIM
    Set5FFCNet22.160.788 922.820.825 623.200.842 6
    G2LNet30.560.898 634.110.943 636.030.959 3
    Set11FFCNet23.980.825 324.290.869 726.720.904 6
    G2LNet27.950.868 731.210.922 533.710.950 0
    BSD68FFCNet22.670.734 623.640.813 425.110.854 1
    G2LNet28.390.789 629.590.882 031.290.918 6
    Table 4. Comparing results with the same structured model without convolutional filter flow
    算法名称平均运行时间(CPU/GPU)
    MH2222.703 468 91/-
    ReconNet23-/0.052 174 221
    CSNet9-/0.049 654 527
    ISTA-Net+24-/0.009 718 182
    G2LNet-/0.123 254 545
    Table 5. Sampling rate is 10%, Comparison of the average running time of different algorithms in Set11
    Yuanhong ZHONG, Qianfeng XU, Yujie ZHOU, Shanshan WANG. Image reconstruction based on deep compressive sensing combined with global and local features[J]. Optics and Precision Engineering, 2023, 31(14): 2135
    Download Citation