Optics and Precision Engineering, Volume. 31, Issue 14, 2135(2023)
Image reconstruction based on deep compressive sensing combined with global and local features
Figures & Tables(11)
Fig. 1. Global-to-Local Compressive Sensing Image Reconstruction Model Structure
Fig. 3. Sampling rate is 10%， and the reconstruction images of each algorithm on the image House are compared
Fig. 4. Sampling rate is 20%， and the reconstruction image comparison of each algorithm on the image Monarch
Fig. 5. Change curve of loss with the number of training iterations （epochs） at 20% sampling rate
Fig. 6. G2LNet reconstruction image and filter flow visualization at 30% sampling rate
Table 1. PSNR of reconstructed images of different algorithms in 8 pictures
View tableView in ArticleTable 1. PSNR of reconstructed images of different algorithms in 8 pictures
采样率 算法 Monarch House Barbara Lena Parrots Peppers Boats C.man 10% MH［ 22 ］23.19 30.30 26.73 26.12 25.35 26.00 26.11 22.12 ReconNet［ 23 ］21.51 26.69 22.50 24.47 23.23 22.67 24.15 21.66 CSNet［ 9 ］26.73 31.68 24.24 28.57 27.40 26.66 28.80 24.92 ISTA-Net+［ 24 ］25.72 30.49 23.52 27.50 26.37 27.13 27.41 23.76 G2LNet 28.07 32.00 25.23 28.86 28.88 27.61 28.95 25.46 20% MH［ 22 ］27.10 33.84 30.81 29.81 29.21 29.78 29.91 25.87 ReconNet［ 23 ］22.89 27.95 22.87 25.39 24.56 24.04 25.98 22.64 CSNet［ 9 ］29.57 33.42 24.98 30.75 29.77 28.42 30.97 26.79 ISTA-Net+［ 24 ］31.01 34.99 26.78 31.14 29.96 30.45 31.91 27.65 G2LNet 31.61 35.16 27.27 31.71 31.12 31.38 32.44 28.30 30% MH［ 22 ］29.20 35.69 32.99 31.99 31.00 31.35 32.25 28.08 ReconNet［ 23 ］29.21 33.61 25.65 33.74 26.88 29.77 30.20 26.90 CSNet［ 9 ］32.58 36.47 28.38 33.35 32.87 30.75 33.21 28.64 ISTA-Net+［ 24 ］34.80 37.07 30.13 33.45 32.91 35.54 35.22 30.35 G2LNet 35.27 37.56 30.87 34.26 33.58 35.81 35.87 31.28
Table 2. PSNR of reconstructed images of different algorithms in multiple datasets
View tableView in ArticleTable 2. PSNR of reconstructed images of different algorithms in multiple datasets
采样率 数据集 MH［ 22 ］ReconNet［ 23 ］CSNet［ 9 ］ISTA-Net+［ 24 ］G2LNet 10% Set5 28.56 24.31 31.54 28.61 30.56 Set11 25.82 22.45 27.37 26.49 27.95 BSD68 23.97 23.62 26.12 25.30 26.66 Avg. 26.12 23.46 28.34 26.80 28.39 20% Set5 26.12 23.46 28.34 26.80 28.39 Set11 28.93 24.44 29.33 30.80 31.21 BSD68 26.98 25.12 29.51 29.03 29.59 Avg. 27.34 24.34 29.06 28.88 29.73 30% Set5 34.09 27.78 36.69 35.45 36.03 Set11 31.55 25.74 30.98 33.70 33.71 BSD68 28.79 26.27 31.01 30.35 31.29 Avg. 31.48 26.60 32.89 33.17 33.68
Table 3. Comparison results of SSIM of reconstructed images with different algorithms
View tableView in ArticleTable 3. Comparison results of SSIM of reconstructed images with different algorithms
采样率 数据集 MH［ 22 ］ReconNet［ 23 ］CSNet［ 9 ］ISTA-Net+［ 24 ］G2LNet 10% Set5 0.835 0 0.734 1 0.901 0 0.839 8 0.898 6 Set11 0.782 2 0.624 5 0.841 2 0.803 6 0.868 7 BSD68 0.655 9 0.648 2 0.772 4 0.700 1 0.789 6 Avg. 0.757 7 0.668 9 0.838 2 0.781 2 0.852 3 20% Set5 0.888 1 0.794 0 0.950 1 0.902 1 0.943 6 Set11 0.872 1 0.732 4 0.901 2 0.913 1 0.922 5 BSD68 0.769 3 0.705 2 0.875 7 0.864 2 0.882 0 Avg. 0.843 2 0.743 9 0.909 0 0.893 1 0.916 0 30% Set5 0.918 5 0.823 3 0.963 3 0.934 1 0.959 3 Set11 0.906 3 0.804 6 0.891 2 0.938 2 0.950 0 BSD68 0.827 6 0.784 2 0.882 4 0.878 2 0.918 6 Avg. 0.884 1 0.804 0 0.912 3 0.916 8 0.942 6
Table 4. Comparing results with the same structured model without convolutional filter flow
View tableView in ArticleTable 4. Comparing results with the same structured model without convolutional filter flow
数据集 算法模型 采样率 10% 20% 30% PSNR SSIM PSNR SSIM PSNR SSIM Set5 FFCNet 22.16 0.788 9 22.82 0.825 6 23.20 0.842 6 G2LNet 30.56 0.898 6 34.11 0.943 6 36.03 0.959 3 Set11 FFCNet 23.98 0.825 3 24.29 0.869 7 26.72 0.904 6 G2LNet 27.95 0.868 7 31.21 0.922 5 33.71 0.950 0 BSD68 FFCNet 22.67 0.734 6 23.64 0.813 4 25.11 0.854 1 G2LNet 28.39 0.789 6 29.59 0.882 0 31.29 0.918 6
Table 5. Sampling rate is 10%， Comparison of the average running time of different algorithms in Set11
View tableView in ArticleTable 5. Sampling rate is 10%， Comparison of the average running time of different algorithms in Set11
算法名称 平均运行时间（CPU/GPU） MH［ 22 ］22.703 468 91/- ReconNet［ 23 ］-/0.052 174 221 CSNet［ 9 ］-/0.049 654 527 ISTA-Net+［ 24 ］-/0.009 718 182 G2LNet -/0.123 254 545
Tools
Get Citation
Copy Citation Text
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
Paper Information
Category: Information Sciences
Received: Dec. 6, 2022
Accepted: --
Published Online: Aug. 2, 2023
The Author Email: ZHONG Yuanhong (zhongyh@cqu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20