Pith. sign in

REVIEW 4 major objections 6 minor 57 references

Advancing Vehicle Plate Recognition: Multitasking Visual Language Models with VehiclePaliGemma

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Fine-tuning a small open VLM on synthetic Malaysian plates lets it beat GPT-4o on real blurry plates.

desk verdict Useful applied VLM paper with an open-sourced model, but the headline superiority claim rests on comparing models under different prompts and needs a fixed-protocol redo. read the letter →

arxiv 2412.14197 v1 pith:4B6ANPPX submitted 2024-12-14 cs.CV cs.LG

Nouar AlDahoul , Myles Joshua Toledo Tan , Raghava Reddy Tera , Hezerul Abdul Karim , Chee How Lim , Manish Kumar Mishra , Yasir Zaki This is my paper · ORCID
classification cs.CVcs.LG
keywords LicensePlateRecognitionVisualLanguageModelsPaliGemmaFine-tuningOpticalCharacterSyntheticdataMalaysianplatesMultitaskingVLM
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that fine-tuning an open 3-billion-parameter visual language model, PaliGemma, on a small set of synthetic Malaysian license plates yields a recognition engine that outperforms much larger commercial VLMs and traditional OCR systems on hard, real-world plate images. On a 258-image set of blurry, low-quality, tightly spaced Malaysian plates, the fine-tuned VehiclePaliGemma reaches 87.6% plate-level accuracy and 97.66% character-level accuracy, compared with 86% for GPT-4o and under 41% for Tesseract, EasyOCR, and KerasOCR. The same model, run through a prompt-driven car and plate detection pipeline, can also localize and read plates in multi-car scenes while filtering by the car's color or model. The paper's point is that a small open VLM fine-tuned on task-specific synthetic data is a practical alternative to closed commercial models for a domain where data are scarce and images are degraded.

What carries the argument

The central object is VehiclePaliGemma, a fine-tuned PaliGemma VLM whose vision encoder is SigLIP-So400m and text decoder is Gemma-2b. The fine-tuning data are 600 synthetic single-line and two-line Malaysian plates in the fixed AAA1234 format, augmented by rotation, blur, Gaussian noise, and salt-and-pepper noise. The fine-tuned model is then prompted with bare instructions such as 'extract the text from the image,' which the paper finds more effective than format-specifying prompts. For multi-car scenes, the machinery is a four-step prompt chain: detect cars with the pre-trained PaliGemma, detect each plate, optionally verify the car's color or model, then extract the plate text.

What would settle it

Run VehiclePaliGemma on a held-out set of Malaysian plates that do not follow the AAA1234 layout, for instance six-character plates such as PJN214 or plates ending in a letter like W1209G, and measure plate-level accuracy; the paper's own examples predict these would mostly fail by inserting or reordering characters, so accuracy near chance on such plates would confirm a format prior rather than genuine reading.

Watch

Extended reading notes

Core claim

VehiclePaliGemma, a PaliGemma-3b model fine-tuned for five epochs on 600 synthetic images of Malaysian plates (three letters followed by four numbers, with rotation, blur, and noise), recognizes complex real Malaysian plates better than all compared alternatives, including GPT-4o, Gemini 1.5 Pro, Claude 3.5 Sonnet, Llama 3.2, LLaVA-NeXT, and VILA. It achieves 87.6% plate-level accuracy (226 of 258 plates) and 97.66% character-level accuracy (1710 of 1751 characters), processing plates at roughly 7 frames per second on an A100 GPU. In multitasking mode, coupling the fine-tuned model with the pre-trained PaliGemma's detection prompts lets the system answer queries like 'extract all characters from the plate of the green Toyota car,' correctly reading all plates in 94.32% of 140 multi-car web images while the GPT-4o-based VehicleGPT counterpart achieves 97.16%.

Load-bearing premise

The paper assumes that fine-tuning on 600 synthetic images of one fixed plate format (three letters then four numbers) teaches the model to recognize real Malaysian plates of any format; the paper's own error analysis shows this assumption fails when plates deviate from that format.

Editorial extensions

If this is right

  • If the result holds, a 3-billion-parameter open model fine-tuned on a few hundred synthetic images can beat closed 100B-plus parameter commercial VLMs on a narrow OCR task, changing the cost calculus for license plate recognition deployments.
  • Traditional OCR pipelines (Tesseract, EasyOCR, KerasOCR) fall below 41% plate-level accuracy on this dataset, while VLM-based approaches, even zero-shot, clear 55% to 86%, suggesting degraded-text OCR benchmarks should be re-baselined against VLMs.
  • Prompt wording is part of the recognition system: the paper shows plate accuracy swings by 4 to 23 percentage points depending on prompt phrasing, so prompt choice is a genuine hyperparameter.
  • A single prompted VLM pipeline can replace separate car detection, plate detection, color recognition, and OCR modules in multi-car scenes, reducing the number of specialized models needed.
  • Synthetic training data with a single fixed plate format is enough to boost plate-level accuracy by 18 percentage points, but the model remains sensitive to format deviations, as the paper's own error cases show.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported accuracy probably overstates genuine generalization because the 600 synthetic images encode a rigid AAA1234 prior; the error cases (W1209G predicted as WI2096, inserted digits on six-character plates) show format bias, so real-world deployment would need more format diversity.
  • Extending the synthetic data to include six-character plates, trailing letters, and Malaysian commercial or special formats would likely close most remaining errors, and this is a cheap, directly testable next step.
  • The multitasking results (94% to 97% image-level accuracy) are not directly comparable to the 87.6% single-plate benchmark because the counting is per-image over 140 scraped web images with an unknown plate distribution.
  • The same fine-tuning recipe with country-specific synthetic plates is a plausible fast path to license plate recognition systems for other nations, provided the format prior is widened.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper evaluates eight visual language models (GPT-4o, Gemini 1.5 Pro/Flash, PaliGemma, Llama 3.2, Claude 3.5 Sonnet, LLaVA-NeXT, VILA, moondream2) and three OCR baselines on a 258-image Malaysian license plate dataset with challenging conditions. The authors introduce VehiclePaliGemma, a PaliGemma 3B model fine-tuned on 600 synthetic images, and report that it achieves 87.6% plate-level accuracy and 97.66% character-level accuracy, outperforming all compared models. A prompt-sensitivity experiment compares four prompts on GPT-4o, Gemini 1.5 Pro, and PaliGemma variants, and a multitasking extension uses prompts referring to car color or model to detect and recognize plates in multi-car images.

Significance. If the central claim were established, the contribution would be practically useful: an open-sourced, fine-tuned 3B VLM that beats much larger commercial VLMs on complex plates, with reported throughput of 7 frames per second on an A100 GPU and a multitasking capability that localizes and reads plates from multi-car scenes. The authors should be credited for releasing the fine-tuned model and for evaluating several open and closed VLMs under a common task. However, the headline superiority claim is currently not supported because the comparison protocol is inconsistent across models and the prompt-sensitivity results in the paper itself show that GPT-4o with prompt3 reaches 88.0%, above the reported 87.6% for VehiclePaliGemma. The small test set (258 images) and the synthetic fine-tuning data with a single plate format further limit the strength of the generalization claims.

major comments (4)
  1. [Section 5.1, Tables 2 and 3] The main comparison in Table 2 is not apples-to-apples. Section 4.2 states that the comparison prompt is 'Extract three letters and four numbers...' (Prompt2), but Section 5.1 reports that for VehiclePaliGemma the plate accuracy with Prompt4 ('extract the text from the image') is 23% better than with Prompt2, and Table 3 lists PaliGemma's Prompt4 accuracy as 69% versus 46.12% with Prompt2. The manuscript does not state which prompt produced VehiclePaliGemma's 87.6% in Table 2. If the headline number was obtained with Prompt4 while the other VLMs were scored with Prompt2, the comparison is invalid. Moreover, Table 3 shows GPT-4o reaches 88.0% with Prompt3, exceeding the reported 87.6%, so the asserted superiority over GPT-4o is contradicted by the paper's own data unless a single consistent protocol is applied.
  2. [Section 5.1 and Section 4.1.1] The prompt-sensitivity analysis selects the best prompt on the same 258-image test set used for the final accuracy claim, without a validation split or any correction for multiple comparisons. Because the paper shows that prompt choice changes accuracy by double-digit percentage points, the reported 87.6% should be treated as a test-set-selected maximum rather than the performance of a fixed evaluation protocol. The authors should either report results for all models under every prompt, or pre-register a single prompt and evaluate all models under that prompt, ideally on a held-out split.
  3. [Section 4.1.2 and Section 5.1 error analysis] Fine-tuning on 600 synthetic images with a fixed format of three letters followed by four numbers introduces a strong format prior, and the paper's own error analysis shows that VehiclePaliGemma inserts characters and reorders letters when plates deviate from that format (e.g., six-character plates and plates ending with a letter). This means the reported accuracy on the 258-image dataset may reflect successful exploitation of the dominant format in that dataset rather than robust recognition of arbitrary Malaysian plates. The authors should report accuracy separately for plates that conform to the synthetic training format and those that do not, and discuss how the format prior affects the claimed generalization.
  4. [Section 5, Tables 1 and 2] The difference between the top models is very small relative to the dataset size: VehiclePaliGemma correctly recognizes 226 plates and GPT-4o 222 plates under the stated comparison, a difference of four plates out of 258, while GPT-4o with Prompt3 recognizes 227. No confidence intervals, statistical tests, or per-prompt variance are reported, so the claimed 'superior performance' is not established even under a consistent protocol. The authors should provide error bars or a significance test, or weaken the claim to a practical equivalence.
minor comments (6)
  1. [Section 5, text near Table 2] The text says EasyOCR predicted correctly 87 images, but Table 2 lists 79 correctly predicted plates; this inconsistency should be corrected.
  2. [Section 5.1, Prompt2 description] The text states that prompt2 specifies 'four letters and three numbers', but the quoted prompt and the actual Malaysian format are three letters and four numbers; the wording should be corrected.
  3. [Section 5.1, Table 3 caption] Table 3 is titled 'Prompt sensitivity in GPT-4o and Gemini 1.5 Pro' but includes rows for Pre-trained PaliGemma with Prompt2 and Prompt4; the caption should be updated to include PaliGemma.
  4. [Section 5.1] There is a typo 'Gemini 1.6 Pro' in the sentence about models that show good performance with prompt2; the intended model is Gemini 1.5 Pro.
  5. [Section 5.2] The text refers to 'PaliGemmaGPT' in the paragraph describing the multitasking evaluation; this should be 'VehiclePaliGemma'.
  6. [Section 5, paragraph after Table 2] The text mentions 'VILA-NEXT 34b' when describing the lowest accuracy among large VLMs, but the table and model names elsewhere refer to 'LLaVA-NeXT-34b'; the naming should be made consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracies are empirical measurements, and the prompt-sensitivity inconsistency in Section 5.1 is a validity concern, not a circular derivation.

full rationale

This paper does not claim a first-principles derivation; its central result (VehiclePaliGemma 87.6% plate accuracy, 97.66% character accuracy) is an observed count on a fixed 258-image test set after fine-tuning on a separate 600-image synthetic set. No equation or construction makes the test accuracy equal to the training data, the prompt, or any fitted parameter. The strongest challenge is in Section 5.1: the paper states the comparison prompt is Prompt2 (Section 4.2), but then reports that VehiclePaliGemma's Prompt4 accuracy was 23 percentage points higher than Prompt2, and Table 3 shows GPT-4o reaching 88% with Prompt3, above the claimed 87.6%. That is a serious benchmarking/reporting inconsistency that weakens the headline comparison, but it does not reduce the reported number to an input by definition: 226/258 is still a measured outcome contingent on model and data. The self-citation [6] for OCR baseline failure is not load-bearing because the current paper re-evaluates those baselines in Table 2. The paper's own limitations in Section 5.1 (P/R confusion, inserted digits, letter-number reordering) are generalization failures, not circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central empirical claim rests on the test set labels, the synthetic-to-real transfer of fine-tuning, and the assumed plate format. No new theoretical entities or forces are introduced; the main hidden choices are the test-set-selected prompt and the fixed synthetic format prior.

free parameters (2)
  • Evaluation prompt for VehiclePaliGemma = prompt4: "extract the text from the image"
    The prompt was chosen after measuring test-set accuracy across options in Section 5.1; it improved accuracy by 23 percentage points over prompt2, and the headline 87.6% result is tied to this test-set-selected prompt.
  • Synthetic training data format prior = 3 letters + 4 numbers, single-line and two-line variants
    Fine-tuning used 600 synthetic images with only this format, which biases the model and causes systematic errors on six-character plates or plates with trailing letters, as documented in Section 5.1.
assumptions (3)
  • domain assumption The 258-image gold set labels are correct and reliable as ground truth.
    Section 4.1.1: labels come from majority voting among three annotators, but no inter-annotator agreement metric is reported; every accuracy figure depends on these labels.
  • domain assumption Fine-tuning on synthetic plate images generalizes to real-world complex plates.
    Sections 4.1.2 and 5: the model was trained only on synthetic images and never validated on real data during training; the observed format-mismatch errors in Section 5.1 show this assumption fails partially.
  • domain assumption The test set consists of plates that fit the assumed 'three letters followed by four numbers' format.
    The prompt and training data assume this format, but the paper's own error analysis in Section 5.1 reveals six-character plates and letter-terminated plates in the test set, violating the assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Vehicle Plate Recognition: Multitasking Visual Language Models with VehiclePaliGemma." pith.science (2026). https://pith.science/paper/4B6ANPPX

@misc{pith2026241214197,
  author       = {Pith},
  title        = {Pith review of: Advancing Vehicle Plate Recognition: Multitasking Visual Language Models with VehiclePaliGemma},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4B6ANPPX}},
  note         = {Machine review of arXiv:2412.14197}
}
read the original abstract

License plate recognition (LPR) involves automated systems that utilize cameras and computer vision to read vehicle license plates. Such plates collected through LPR can then be compared against databases to identify stolen vehicles, uninsured drivers, crime suspects, and more. The LPR system plays a significant role in saving time for institutions such as the police force. In the past, LPR relied heavily on Optical Character Recognition (OCR), which has been widely explored to recognize characters in images. Usually, collected plate images suffer from various limitations, including noise, blurring, weather conditions, and close characters, making the recognition complex. Existing LPR methods still require significant improvement, especially for distorted images. To fill this gap, we propose utilizing visual language models (VLMs) such as OpenAI GPT4o, Google Gemini 1.5, Google PaliGemma (Pathways Language and Image model + Gemma model), Meta Llama 3.2, Anthropic Claude 3.5 Sonnet, LLaVA, NVIDIA VILA, and moondream2 to recognize such unclear plates with close characters. This paper evaluates the VLM's capability to address the aforementioned problems. Additionally, we introduce ``VehiclePaliGemma'', a fine-tuned Open-sourced PaliGemma VLM designed to recognize plates under challenging conditions. We compared our proposed VehiclePaliGemma with state-of-the-art methods and other VLMs using a dataset of Malaysian license plates collected under complex conditions. The results indicate that VehiclePaliGemma achieved superior performance with an accuracy of 87.6\%. Moreover, it is able to predict the car's plate at a speed of 7 frames per second using A100-80GB GPU. Finally, we explored the multitasking capability of VehiclePaliGemma model to accurately identify plates containing multiple cars of various models and colors, with plates positioned and oriented in different directions.

Figures

Figures reproduced from arXiv: 2412.14197 by the authors.

Figure 1
Figure 1. Sample complex license plates from the used dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. The proposed solution block diagram [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Character-level accuracy heatmaps for different vision models. [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: VehiclePaliGemma’s block diagram [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]
Figure 5
Figure 5. Figure 5: samples of cars detected by VehiclePaliGemma [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 6
Figure 6. Figure 6: samples of license plates detected by VehiclePaliGemma [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Example for prompt 1 featuring red cars. [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Example for prompt 2 featuring a BMW blue car. [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: Example for prompt 1 and prompt 3 featuring a red car and a PERODUA car, [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 48 canonical work pages

  1. [1]

    E., Anagnostopoulos, I

    Anagnostopoulos, C.-N. E., Anagnostopoulos, I. E., Psoroulas, I. D., Loumos, V. & Kayafas, E. License plate recognition from still images 27 and video sequences: A survey. IEEE Transactions on intelligent trans- portation systems 9, 377–391 (2008)

    work page 2008

  2. [2]

    & Shah, S

    Lubna, Mufti, N. & Shah, S. A. A. Automatic number plate recognition: A detailed survey of relevant algorithms. Sensors 21, 3028 (2021)

    work page 2021

  3. [3]

    WIJERS, P. J. Implementing automated enforcement in emerging economies. 17th International Road Federation World Meeting(2013)

    work page 2013

  4. [4]

    & Badawy, W

    Du, S., Ibrahim, M., Shehata, M. & Badawy, W. Automatic license plate recognition (alpr): A state-of-the-art review. IEEE Transactions on circuits and systems for video technology23, 311–325 (2012)

    work page 2012

  5. [5]

    & Nasir, N

    Kamaruzaman, M. & Nasir, N. R. M. Parkey: ticket-less parking sys- tem using license plate recognition approach. In Journal of Physics: Conference Series, vol. 1860, 012006 (IOP Publishing, 2021)

    work page 2021

  6. [6]

    A., Shahid, R

    Idrose, H., AlDahoul, N., Karim, H. A., Shahid, R. & Mishra, M. K. An evaluation of various pre-trained optical character recognition mod- els for complex license plates. In Multimedia University Engineering Conference (MECON 2022), 21–27 (Atlantis Press, 2022)

    work page 2022

  7. [7]

    Y., Adrio, K., Tanuwijaya, K

    Sugiyono, A. Y., Adrio, K., Tanuwijaya, K. & Suryaningrum, K. M. Ex- tracting information from vehicle registration plate using ocr tesseract. Procedia Computer Science227, 932–938 (2023)

    work page 2023

  8. [8]

    Bishop, C. M. & Nasrabadi, N. M. Pattern recognition and machine learning, vol. 4 (Springer, 2006)

    work page 2006

Show all 57 references
  1. [9]

    & Vincent, P

    Bengio, Y., Ducharme, R. & Vincent, P. A neural probabilistic language model. Advances in neural information processing systems13 (2000)

    2000

  2. [10]

    Yang, X. et al. A large language model for electronic health records. NPJ digital medicine5, 194 (2022)

    2022

  3. [11]

    C., Ursut , iu, D., Samoila, C

    Modran, H., Bogdan, I. C., Ursut , iu, D., Samoila, C. & Modran, P. L. Llm intelligent agent tutoring in higher education courses using a rag approach. Preprints (2024)

    2024

  4. [12]

    Vaswani, A. et al. Attention is all you need. Advances in neural infor- mation processing systems30 (2017). 28

    2017

  5. [13]

    & Toutanova, K

    Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  6. [14]

    https://openai.com/index/hello-gpt-4o// (2024)

    Hello gpt-4o. https://openai.com/index/hello-gpt-4o// (2024)

    2024

  7. [15]

    Radford, A. et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748–8763 (PMLR, 2021)

    2021

  8. [16]

    https://blog

    Gpt-4o: The comprehensive guide and explanation. https://blog. roboflow.com/gpt-4o-vision-use-cases/ (2024)

    2024

  9. [17]

    Gemini 1.5 technical report

    Gemini Team, G. Gemini 1.5 technical report. https://storage. googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf (2024)

    2024

  10. [18]

    https://blog.google/technology/ai/ google-gemini-next-generation-model-february-2024/ #architecture (2024)

    Introducing gemini 1.5, google’s next-generation ai model. https://blog.google/technology/ai/ google-gemini-next-generation-model-february-2024/ #architecture (2024)

    2024

  11. [19]

    Beyer, L. et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726 (2024)

    2024 arXiv

  12. [20]

    meta-llama/llama-3.2-11b-vision-instruct

    Face, H. meta-llama/llama-3.2-11b-vision-instruct. https:// huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct

  13. [21]

    Claude 3.5 sonnet

    Anthropic. Claude 3.5 sonnet. https://www.anthropic.com/news/ claude-3-5-sonnet

  14. [22]

    https://encord

    Llava: Large language and vision assistant explained. https://encord. com/blog/llava-large-language-vision-assistant/ (2024)

    2024

  15. [23]

    Liu, H. et al. Llava-next: Improved reasoning, ocr, and world knowledge (2024). URL https://llava-vl.github.io/blog/ 2024-01-30-llava-next/

    2024

  16. [24]

    Multimodal (visual and language) understanding with llava- next

    Dang, P. Multimodal (visual and language) understanding with llava- next. https://rocm.blogs.amd.com/artificial-intelligence/ llava-next/README.html (2023). 29

    2023

  17. [25]

    Lin, J. et al. Vila: On pre-training for visual language models. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 26689–26699 (2024)

    2024

  18. [26]

    https://huggingface.co/vikhyatk/moondream2 (2024)

    moondream2. https://huggingface.co/vikhyatk/moondream2 (2024)

    2024

  19. [27]

    https://moondream.ai/ (2024)

    Vision ai for devs. https://moondream.ai/ (2024)

    2024

  20. [28]

    & Rahmat, R

    Gunawan, D., Rohimah, W. & Rahmat, R. Automatic number plate recognition for indonesian license plate by using k-nearest neighbor al- gorithm. In IOP conference series: materials science and engineering, vol. 648, 012011 (IOP Publishing, 2019)

    2019

  21. [29]

    Canny edge-detection based vehicle plate recognition

    Mousa, A. Canny edge-detection based vehicle plate recognition. In- ternational Journal of signal processing, Image processing and pattern recognition 5, 1–8 (2012)

    2012

  22. [30]

    & Suruthi, M

    Aruna, V., Ravi, S. & Suruthi, M. Detection and recognition of li- cense plates using color image processing. In International Confer- ence on Communications and Cyber Physical Engineering 2018, 133–140 (Springer, 2024)

    2018

  23. [31]

    & Hinton, G

    Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. Advances in neural informa- tion processing systems25 (2012)

    2012

  24. [32]

    & Hinton, G

    LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. nature 521, 436–444 (2015)

    2015

  25. [33]

    & Jung, C

    Montazzolli, S. & Jung, C. Real-time brazilian license plate detection and recognition using deep convolutional neural networks. In 2017 30th SIBGRAPI conference on graphics, patterns and images (SIBGRAPI), 55–62 (IEEE, 2017)

    2017

  26. [34]

    & Huang, T

    He, H., He, S. & Huang, T. License plate recognition based on three different neural networks. In 2022 IEEE 4th international conference on civil aviation safety and information technology (ICCASIT), 215–220 (IEEE, 2022). 30

    2022

  27. [35]

    & Zisserman, A

    Simonyan, K. & Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556(2014)

    2014 arXiv

  28. [36]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770–778 (2016)

    2016

  29. [37]

    & Malik, J

    Girshick, R., Donahue, J., Darrell, T. & Malik, J. Rich feature hierar- chies for accurate object detection and semantic segmentation. In Pro- ceedings of the IEEE conference on computer vision and pattern recog- nition, 580–587 (2014)

    2014

  30. [38]

    Fast r-cnn

    Girshick, R. Fast r-cnn. In Proceedings of the IEEE international con- ference on computer vision, 1440–1448 (2015)

    2015

  31. [39]

    & Sun, J

    Ren, S., He, K., Girshick, R. & Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems28 (2015)

    2015

  32. [40]

    & Touati, Y

    Saidani, T. & Touati, Y. E. A vehicle plate recognition system based on deep learning algorithms. Multimedia Tools and Applications 80, 36237–36248 (2021)

    2021

  33. [41]

    & Farhadi, A

    Redmon, J., Divvala, S., Girshick, R. & Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 779–788 (2016)

    2016

  34. [42]

    & Farhadi, A

    Redmon, J. & Farhadi, A. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)

    2018 arXiv

  35. [43]

    Hendryli, J., Herwindiati, D. E. et al. Automatic license plate recog- nition for parking system using convolutional neural networks. In 2020 International Conference on Information Management and Technology (ICIMTech), 71–74 (IEEE, 2020)

    2020

  36. [44]

    Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929(2020)

    2020 arXiv

  37. [45]

    & Jia, W

    Zhang, T. & Jia, W. Automatic license plate recognition using trans- former. In Fourteenth International Conference on Graphics and Image Processing (ICGIP 2022), vol. 12705, 129–138 (SPIE, 2023). 31

    2022

  38. [46]

    Abdelhamed, A., Afifi, M. & Go, A. What do you see? enhancing zero- shot image classification with multimodal large language models. arXiv preprint arXiv:2405.15668 (2024)

    2024 arXiv

  39. [47]

    Ramesh, A. et al. Zero-shot text-to-image generation. In Meila, M. & Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, vol. 139 of Proceedings of Machine Learning Re- search, 8821–8831 (PMLR, 2021). URL https://proceedings.mlr. press/v139/r...

    2021

  40. [48]

    Antol, S. et al. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, 2425–2433 (2015)

    2015

  41. [49]

    https://tesseract-ocr.github.io/ (2022)

    Tesseract documentation. https://tesseract-ocr.github.io/ (2022)

    2022

  42. [50]

    https://github.com/JaidedAI/EasyOCR (2022)

    Easyocr. https://github.com/JaidedAI/EasyOCR (2022)

    2022

  43. [51]

    https://github.com/faustomorales/keras-ocr (2022)

    keras-ocr. https://github.com/faustomorales/keras-ocr (2022)

    2022

  44. [52]

    https://gotapway.com/solutions/vehicletrack (2024)

    Vehicletrack. https://gotapway.com/solutions/vehicletrack (2024)

    2024

  45. [53]

    meta-llama/llama-3.1-8b-instruct

    Face, H. meta-llama/llama-3.1-8b-instruct. https://huggingface.co/ meta-llama/Llama-3.1-8B-Instruct

  46. [54]

    Brown, T. et al. Language models are few-shot learners. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M. & Lin, H. (eds.) Advances in Neural Information Processing Systems, vol. 33, 1877–1901 (Curran Associates, Inc., 2020). URL https://proceedings.neurips.cc/paper_file...

    1901

  47. [55]

    https://nanonets.com/blog/ ocr-with-tesseract/ (2024)

    Ocr using pytesseract and opencv. https://nanonets.com/blog/ ocr-with-tesseract/ (2024)

    2024

  48. [56]

    & Midina, S

    Smelyakov, K., Chupryna, A., Darahan, D. & Midina, S. Effectiveness of modern text recognition solutions and tools for common data sources. In COLINS, 154–165 (2021). 32

    2021

  49. [57]

    & Arun, D

    Vedhaviyassh, D., Sudhan, R., Saranya, G., Safa, M. & Arun, D. Com- parative analysis of easyocr and tesseractocr for automatic license plate recognition using deep learning algorithm. In 2022 6th International Conference on Electronics, Communication and Aerospace Technology,...

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.