Author Contributions
Conceptualization, S.B.Y.; methodology, S.-J.L. and S.B.Y.; software, S.-J.L.; validation, S.-J.L.; formal analysis, S.-J.L. and S.B.Y.; investigation, S.-J.L. and S.B.Y.; resources, S.-J.L. and S.B.Y.; data curation, S.-J.L. and S.B.Y.; writing—original draft preparation, S.-J.L. and S.B.Y.; writing—review and editing, S.-J.L. and S.B.Y.; visualization, S.B.Y.; supervision, S.B.Y.; project administration, S.B.Y.; funding acquisition, S.B.Y. Both authors have read and agreed to the published version of the manuscript.
Figure 1.
Iterative up and down blocks for SR framework. Two characters were masked due to privacy policies.
Figure 1.
Iterative up and down blocks for SR framework. Two characters were masked due to privacy policies.
Figure 2.
Trends of the average PSNR and SSIM results, according to the checkpoint.
Figure 2.
Trends of the average PSNR and SSIM results, according to the checkpoint.
Figure 3.
Proposed loss function for the SR framework. Two characters were masked due to privacy policies.
Figure 3.
Proposed loss function for the SR framework. Two characters were masked due to privacy policies.
Figure 4.
Examples of artificial license plate images.
Figure 4.
Examples of artificial license plate images.
Figure 5.
Flowchart of the integrated SR and character recognition. Two characters were masked due to privacy policies.
Figure 5.
Flowchart of the integrated SR and character recognition. Two characters were masked due to privacy policies.
Figure 6.
SR results of the first plate image (4×). (a) HR (128 × 64), (b) Bicubic, (c) DRN, (d) MZSR, (e) USRNet, (f) DBPN, (g) Proposed. Two characters were masked due to privacy policies.
Figure 6.
SR results of the first plate image (4×). (a) HR (128 × 64), (b) Bicubic, (c) DRN, (d) MZSR, (e) USRNet, (f) DBPN, (g) Proposed. Two characters were masked due to privacy policies.
Figure 7.
SR results of the second plate image (4×). (a) HR (140 × 48), (b) Bicubic, (c) DRN, (d) MZSR, (e) USRNet, (f) DBPN, (g) Proposed. Two characters were masked due to privacy policies.
Figure 7.
SR results of the second plate image (4×). (a) HR (140 × 48), (b) Bicubic, (c) DRN, (d) MZSR, (e) USRNet, (f) DBPN, (g) Proposed. Two characters were masked due to privacy policies.
Figure 8.
SR results of the third plate image (4×). (a) HR (144 × 40), (b) Bicubic, (c) DRN, (d) MZSR, (e) USRNet, (f) DBPN, (g) Proposed. Two characters were masked due to privacy policies.
Figure 8.
SR results of the third plate image (4×). (a) HR (144 × 40), (b) Bicubic, (c) DRN, (d) MZSR, (e) USRNet, (f) DBPN, (g) Proposed. Two characters were masked due to privacy policies.
Figure 9.
Results of character recognition (a) without and (b) with artificial license plate generation.
Figure 9.
Results of character recognition (a) without and (b) with artificial license plate generation.
Figure 10.
Examples of plate detection results. Two characters were masked due to privacy policies.
Figure 10.
Examples of plate detection results. Two characters were masked due to privacy policies.
Figure 11.
Examples of character recognition results. Two characters were masked due to privacy policies.
Figure 11.
Examples of character recognition results. Two characters were masked due to privacy policies.
Figure 12.
Examples of character recognition results. Recognition results on (a) bicubic results (3×), (b) proposed SR results (3×), and (c) HR images. Two characters were masked due to privacy policies.
Figure 12.
Examples of character recognition results. Recognition results on (a) bicubic results (3×), (b) proposed SR results (3×), and (c) HR images. Two characters were masked due to privacy policies.
Figure 13.
Examples of character recognition results. Recognition results on (a) bicubic results (4×), (b) proposed SR results (4×), and (c) HR images. Two characters were masked due to privacy policies.
Figure 13.
Examples of character recognition results. Recognition results on (a) bicubic results (4×), (b) proposed SR results (4×), and (c) HR images. Two characters were masked due to privacy policies.
Figure 14.
Failure examples of character recognition results. (a) Ground-truth labels on HR images and (b) the proposed SR (4×) and character recognition results obtained from LR images. Two characters were masked due to privacy policies.
Figure 14.
Failure examples of character recognition results. (a) Ground-truth labels on HR images and (b) the proposed SR (4×) and character recognition results obtained from LR images. Two characters were masked due to privacy policies.
Figure 15.
mAP comparison results. (a–c) represent the mAP results on bicubic results (3×), proposed SR results (3×), and HR images, respectively.
Figure 15.
mAP comparison results. (a–c) represent the mAP results on bicubic results (3×), proposed SR results (3×), and HR images, respectively.
Figure 16.
mAP comparison results for only numbers (0–9). (a–c) represent the mAP results on bicubic results (3×), proposed SR results (3×), and HR images, respectively.
Figure 16.
mAP comparison results for only numbers (0–9). (a–c) represent the mAP results on bicubic results (3×), proposed SR results (3×), and HR images, respectively.
Figure 17.
mAP comparison results. (a–d) represent the mAP results on bicubic results (×4), DBPN results (×4), proposed SR results (×4), and HR images, respectively.
Figure 17.
mAP comparison results. (a–d) represent the mAP results on bicubic results (×4), DBPN results (×4), proposed SR results (×4), and HR images, respectively.
Figure 18.
mAP comparison results for only numbers (0–9). (a–d) represent the mAP results on bicubic results (×4), DBPN results (×4), proposed SR results (×4), and HR images, respectively.
Figure 18.
mAP comparison results for only numbers (0–9). (a–d) represent the mAP results on bicubic results (×4), DBPN results (×4), proposed SR results (×4), and HR images, respectively.
Table 1.
Hyper parameters of the SR framework.
Table 1.
Hyper parameters of the SR framework.
Scale | 2× | 3× | 4× |
---|
Kernel size | 6 | 5 | 8 |
Stride | 2 | 3 | 4 |
Padding | 2 | 1 | 2 |
Table 2.
122 classes for character recognition.
Table 2.
122 classes for character recognition.
Class | Character | Class | Character | Class | Character | Class | Character |
---|
0 | 0 | 31 | 강원 | 62 | 두 | 93 | 차 |
1 | 1 | 32 | 경기 | 63 | 드 | 94 | 처 |
2 | 2 | 33 | 경남 | 64 | 라 | 95 | 초 |
3 | 3 | 34 | 경북 | 65 | 러 | 96 | 추 |
4 | 4 | 35 | 광주 | 66 | 로 | 97 | 츠 |
5 | 5 | 36 | 대구 | 67 | 루 | 98 | 카 |
6 | 6 | 37 | 대전 | 68 | 르 | 99 | 커 |
7 | 7 | 38 | 부산 | 69 | 마 | 100 | 코 |
8 | 8 | 39 | 서울 | 70 | 머 | 101 | 쿠 |
9 | 9 | 40 | 세종 | 71 | 모 | 102 | 크 |
10 | 강 | 41 | 울산 | 72 | 무 | 103 | 타 |
11 | 경 | 42 | 인천 | 73 | 므 | 104 | 터 |
12 | 광 | 43 | 전남 | 74 | 바 | 105 | 토 |
13 | 대 | 44 | 전북 | 75 | 배 | 106 | 투 |
14 | 부 | 45 | 제주 | 76 | 버 | 107 | 트 |
15 | 서 | 46 | 충남 | 77 | 보 | 108 | 파 |
16 | 세 | 47 | 충북 | 78 | 브 | 109 | 퍼 |
17 | 울 | 48 | 가 | 79 | 사 | 110 | 포 |
18 | 인 | 49 | 거 | 80 | 소 | 111 | 푸 |
19 | 전 | 50 | 고 | 81 | 수 | 112 | 하 |
20 | 제 | 51 | 공 | 82 | 스 | 113 | 프 |
21 | 충 | 52 | 그 | 83 | 아 | 114 | 호 |
22 | 구 | 53 | 국 | 84 | 어 | 115 | 허 |
23 | 기 | 54 | 나 | 85 | 오 | 116 | 해 |
24 | 남 | 55 | 너 | 86 | 우 | 117 | 후 |
25 | 북 | 56 | 노 | 87 | 육 | 118 | 흐 |
26 | 산 | 57 | 누 | 88 | 자 | 119 | 합 |
27 | 원 | 58 | 느 | 89 | 저 | 120 | 영 |
28 | 종 | 59 | 다 | 90 | 조 | 121 | - |
29 | 주 | 60 | 더 | 91 | 주 | | |
30 | 천 | 61 | 도 | 92 | 즈 | | |
Table 3.
Average PSNR/SSIM of license plates for validation.
Table 3.
Average PSNR/SSIM of license plates for validation.
Scale | 2× | 3× | 4× |
---|
Bicubic | 30.85 | 0.9297 | 26.09 | 0.7960 | 23.68 | 0.6516 |
MZSR | 21.15 | 0.7056 | 21.27 | 0.5942 | 19.53 | 0.3788 |
DRN | - | - | - | - | 25.80 | 0.6945 |
DBPN | 33.72 | 0.9616 | - | - | 25.41 | 0.7739 |
Proposed | 33.75 | 0.9632 | 30.09 | 0.9181 | 26.99 | 0.8430 |
Table 4.
Average mAP (%) comparison of LR license plates for validation of 122 classes (0–121).
Table 4.
Average mAP (%) comparison of LR license plates for validation of 122 classes (0–121).
Scale | 2× | 3× | 4× |
---|
Bicubic | 62.61 | 22.00 | 4.83 |
MZSR | 18.24 | 18.53 | 7.65 |
DRN | - | - | 14.60 |
DBPN | 57.49 | - | 13.58 |
Proposed | 64.78 | 35.56 | 25.13 |
Table 5.
Average mAP (%) comparison of LR license plates for validation of 10 classes (0–9).
Table 5.
Average mAP (%) comparison of LR license plates for validation of 10 classes (0–9).
Scale | 2× | 3× | 4× |
---|
Bicubic | 85.13 | 43.90 | 11.36 |
MZSR | 58.37 | 58.86 | 30.86 |
DRN | - | - | 48.40 |
DBPN | 85.85 | - | 19.80 |
Proposed | 91.70 | 71.80 | 61.30 |