Laser & Optoelectronics Progress, Volume. 59, Issue 24, 2411001(2022)
Blind Restoration Method for Incomplete and Sparse Text Images Based on Content Style Transfer
Figures & Tables(17)
Fig. 2. Structure diagram of blind restoration network for incomplete sparse text image
Fig. 6. Discriminator network structure based on improved self-attention mechanism
Fig. 9. Some regular script data of Ouyang xun, Wang Xizhi, Su Shi, Huang Tingjian, etc
Fig. 10. Comparison of experimental results. (a) Broken images; (b) CycleGAN; (c) CycleGAN (LS+SA); (d) proposed algorithm;(e) original images
Fig. 11. “Xi” “Du” “Nian” and “Huai” calligraphy image AHE. (a) “Xi”calligraphy image AHE; (b) “Du” calligraphy image AHE; (c) “Nian” calligraphy image AHE; (d) “Huai” calligraphy image AHE
Fig. 12. Image preprocessing results. (a) Original images; (b) AHE+SR; (c) text extraction
Fig. 13. Comparison results of different repair algorithms.(a) Adding mask artificially; (b) RFR; (c) proposed algorithm
Table 1. PSNR, SSIM, and RMSE comparison results
View tableTable 1. PSNR, SSIM, and RMSE comparison results
Evaluation index Algorithm Calligraphy image PSNR CycleGAN 12.29 11.98 12.99 14.05 12.14 12.44 13.31 13.32 CycleGAN（LS+SA） 12.63 13.16 13.29 15.23 12.18 13.21 13.59 13.52 Proposed algorithm 13.24 12.41 14.04 15.65 12.73 14.27 14.59 14.71 SSIM CycleGAN 0.7535 0.7374 0.7873 0.8038 0.7252 0.7795 0.8155 0.7854 CycleGAN（LS+SA） 0.7405 0.7669 0.7820 0.8229 0.7308 0.7933 0.8126 0.7817 Proposed algorithm 0.7720 0.7392 0.8014 0.8372 0.7457 0.8144 0.8258 0.8143 RMSE CycleGAN 61.92 64.22 57.14 50.62 62.31 60.87 55.06 55.03 CycleGAN（LS+SA） 59.56 56.02 55.20 44.14 62.69 55.73 53.31 53.75 Proposed algorithm 55.58 61.08 50.64 42.05 58.89 49.30 47.53 46.86
Table 2. Comparison results of PSNR, SSIM, and RMSE under various failure ratios
View tableTable 2. Comparison results of PSNR, SSIM, and RMSE under various failure ratios
Dataset Calligraphy image dataset Mask ratio /% 11.53 12.54 23.73 28.27 32.18 43.39 54.95 62.73 PSNR CycleGAN 13.97 14.53 13.54 14.43 14.11 11.85 11.56 10.11 CycleGAN（LS+SA） 14.40 15.10 14.47 14.54 15.20 12.45 11.74 11.16 Proposed algorithm 14.20 13.75 14.29 14.67 15.63 13.09 12.58 11.81 SSIM CycleGAN 0.7876 0.7968 0.8009 0.7966 0.8395 0.7562 0.7564 0.6988 CycleGAN（LS+SA） 0.8103 0.8380 0.8302 0.8101 0.8639 0.7697 0.7509 0.7162 Proposed algorithm 0.8120 0.7990 0.8254 0.8066 0.8478 0.7921 0.7713 0.7409 RMSE CycleGAN 51.12 47.87 53.67 48.42 50.22 65.14 67.35 79.66 CycleGAN（LA+SA） 48.58 44.82 48.18 47.83 44.31 60.82 66.02 70.51 Proposed algorithm 49.69 52.39 49.24 47.16 42.19 56.52 59.90 65.49
Table 3. CCIR500-1 Subjective evaluation scale
View tableTable 3. CCIR500-1 Subjective evaluation scale
Score Obstruction criterion Quality standard 5 No deterioration in image quality Very good 4 The image quality has deteriorated，not hinder viewing Good 3 The deterioration of image quality slightly hinders viewing General 2 Obstruction to viewing Bad 1 Very serious obstruction of viewing Very bad
Table 4. Subjective evaluation results
View tableTable 4. Subjective evaluation results
Proposed result RFR result Proposed score RFR score 4.2 2.6 3.9 1.7 4.9 2.5 
3.8 2.0
Tools
Get Citation
Copy Citation Text
Dongbin Liu, Huiqin Wang, Ke Wang, Zhan Wang, Gang Zhen. Blind Restoration Method for Incomplete and Sparse Text Images Based on Content Style Transfer[J]. Laser & Optoelectronics Progress, 2022, 59(24): 2411001
Paper Information
Category: Imaging Systems
Received: Sep. 1, 2021
Accepted: Oct. 27, 2021
Published Online: Jan. 11, 2023
The Author Email: Wang Huiqin (hqwang@xauat.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20