Optics and Precision Engineering, Volume. 32, Issue 4, 622(2024)
Multi-spectral image compression by fusing multi-scale feature convolutional neural networks
Figures & Tables(18)
Fig. 1. Multi-scale spatial inter-spectral feature extraction network structure
Table 1. PSNR metrics for MSSA modules with different convolutional kernel sizes
View tableView in ArticleTable 1. PSNR metrics for MSSA modules with different convolutional kernel sizes
Spectral Spacial （5，7） （5，9） （7，9） （5，7） 47.86 47.65 47.49 （5，9） 47.78 47.45 47.62 （7，9） 47.21 47.36 47.55
Table 2. MSA metrics for MSSA modules with different convolutional kernel sizes
View tableView in ArticleTable 2. MSA metrics for MSSA modules with different convolutional kernel sizes
Spectral Special （5，7） （5，9） （7，9） （5，7） 15.5 16.5 15.9 （5，9） 15.8 16.2 16.0 （7，9） 16.7 16.5 16.1
Table 3. PSNR metrics and MSA metrics for SSC modules with different numbers of SSC modules
View tableView in ArticleTable 3. PSNR metrics and MSA metrics for SSC modules with different numbers of SSC modules
Performance number PSNR/dB MSA/（×10-3 rad） K=2 47.34 15.9 K=3 48.11 15.4 K=4 47.92 15.7
Table 4. PSNR， MSA metrics under 7-band test set
View tableView in ArticleTable 4. PSNR， MSA metrics under 7-band test set
Method Bit rate PSNR/dB MSA/（×10-3 rad） Method Bit rate PSNR/dB MSA/（×10-3 rad） Scale-only［ 13 ］0.17 45.59 43.2 Joint［ 14 ］0.16 45.89 39.8 0.27 46.37 37.2 0.25 46.32 31.3 0.37 46.65 29.3 0.31 47.14 28.3 0.42 47.49 24.8 0.46 47.49 21.6 0.52 47.82 22.2 0.55 47.96 19.5 Mean&Scale［ 13 ］0.12 44.13 57.8 3D-SPIHT［ 9 ］0.14 38.25 62.9 0.23 45.85 42.0 0.25 39.6 48.7 0.32 46.34 34.3 0.33 40.38 40.9 0.43 47.16 21.6 0.44 41.92 31.6 0.52 47.28 19.5 0.50 43.1 23.5 JPEG2000［ 8 ］0.12 36.32 58.8 本文方法 0.14 46.4 37.4 0.23 37.6 52.9 0.25 47.08 28.2 0.32 38.31 47.0 0.35 47.6 21.2 0.43 39.82 36.9 0.44 48.09 16.1 0.52 40.72 34.8 0.53 48.73 15.2
Table 5. PSNR， MSA metrics under 8-band test set
View tableView in ArticleTable 5. PSNR， MSA metrics under 8-band test set
Method Bit rate PSNR/dB MSA/（×10-3 rad） Method Bit rate PSNR/dB MSA/（×10-3 rad） Scale-only［ 13 ］0.11 59.95 18.8 Joint［ 14 ］0.12 60.18 16.7 0.19 60.71 17.4 0.23 61.60 15.7 0.32 62.02 16.1 0.37 62.56 15.3 0.42 62.26 15.3 0.46 62.74 15.2 0.55 62.30 15.0 0.55 62.89 15.1 Mean&Scale［ 13 ］0.14 54.08 17.3 3D-SPIHT［ 9 ］0.12 37.34 21.4 0.25 58.53 16.5 0.23 38.74 20.1 0.33 62.22 15.4 0.37 40.23 18.9 0.44 62.80 14.9 0.46 40.74 17.6 0.50 62.96 14.5 0.55 41.34 16.4 JPEG2000［ 8 ］0.11 31.84 46.2 本文方法 0.12 60.63 16.3 0.19 32.46 45.0 0.24 61.84 15.3 0.32 33.14 43.5 0.35 62.91 14.8 0.42 33.95 40.5 0.46 63.28 14.3 0.55 34.19 39.3 0.53 63.41 14.1
Table 6. Codec time and computation for each algorithm
View tableView in ArticleTable 6. Codec time and computation for each algorithm
Methods Enc time
/ms
Dec time
/ms
FLOPs
/G
本文方法 62.7 39.1 337.8 Scale-only［ 13 ］36.5 35.4 54.1 Mean & Scale［ 13 ］37.2 36.8 55.1 Joint［ 14 ］1 755.8 5 986.4 110.8
Tools
Get Citation
Copy Citation Text
Lili ZHANG, Zikun CHEN, Tianpeng PAN, Lele QU. Multi-spectral image compression by fusing multi-scale feature convolutional neural networks[J]. Optics and Precision Engineering, 2024, 32(4): 622
Paper Information
Category:
Received: Jul. 20, 2023
Accepted: --
Published Online: Apr. 2, 2024
The Author Email: CHEN Zikun (chenzikun1@stu.sau.edu.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20