TransLPRNet: Lite Vision-Language Network for Single/Dual-line Chinese License Plate Recognition

Bangjun Lei; Guangzhu Xu; Pengcheng Zuo; Zhi Ke

REVIEW 4 major objections 5 minor 29 references

TransLPRNet: Lite Vision-Language Network for Single/Dual-line Chinese License Plate Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A lightweight vision-language Transformer can recognize single- and double-line Chinese license plates at around 99% accuracy, with a weakly supervised perspective-correction network handling tilted images.

desk verdict A solid engineering paper with a genuine weak-supervision idea for perspective correction, but the double-line results are synthetic-only and the abstract numbers diverge from the body, so the practical claims need tightening. read the letter →

arxiv 2507.17335 v2 pith:GLLQVNWS submitted 2025-07-23 cs.CV cs.CL

Guangzhu Xu , Zhi Ke , Pengcheng Zuo , Bangjun Lei This is my paper

classification cs.CVcs.CL

keywords licenseplaterecognitionChineseplatesdouble-linevision-languagemodelTransformerperspectivecorrectionsyntheticdataweaksupervision

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

TransLPRNet tries to establish that Chinese license plate recognition, including the double-line plates used on trucks and buses, can be handled by one lightweight vision-language Transformer rather than by CNN/CRNN pipelines that flatten characters into a one-dimensional sequence; this matters for real-time parking, toll, and traffic applications that must recognize both plate styles. The proposed network pairs a MobileViTv3 visual encoder with a four-layer text decoder that reads the whole plate as a sequence of image tokens, which lets it model two lines of characters as a single spatial layout. To deal with angled captures, the paper adds a Perspective Transformation Network (PTN) that regresses the plate's four corners and is trained with weak supervision from a front-view/no-front-view classifier, avoiding expensive corner annotation and STN-style training instability. Because real double-line plate images are scarce, the authors synthesize double-line plates, blur them, and paste them onto real CCPD backgrounds, mixing them with real images to build a single/double-line dataset. The paper reports 98.75% average accuracy under coarse localization perturbation and 99.03% under fine perturbation on the corrected mixed test set, at 46 FPS with 5.94M parameters.

What carries the argument

The load-bearing machinery is the encoder-adapter-decoder pipeline plus two supporting mechanisms. The encoder stacks inverted residual blocks and MobileViTv3 blocks: inverted residuals extract local character detail cheaply, while MobileViT blocks apply windowed self-attention to local windows, giving global layout modeling within a small computational budget; the adapter reduces the 7×7×320 feature map to 49 tokens of width 128, and a four-layer transformer text decoder with masked self-attention and cross-attention generates the character sequence autoregressively. The second mechanism, PTN, replaces direct regression of a perspective matrix with regression of four plate-corner coordinates, then computes the 3×3 perspective transform by solving the standard 8-parameter linear system; supervision comes from a MobileNetV3 binary classifier that decides whether an image is a frontal view, so corner labels themselves are not needed for PTN training. The third mechanism is data: program-generated yellow and green double-line plate images are blurred, pasted onto redundant real CCPD backgrounds via perspective transform, and cropped with perturbed boxes, producing a mixed single/double-line dataset while redundancy compression keeps total dataset size unchanged.

What would settle it

Train and evaluate the described pipeline on a held-out set of real double-line Chinese plates photographed from trucks and buses, with no overlap with the synthetic generation templates; if accuracy falls materially below the reported ~99% Base-d figure, the synthetic-data representativeness assumption is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a single lightweight encoder-decoder can unify single- and double-line Chinese license plate recognition. The visual encoder, built from MobileViTv3's inverted residual blocks and windowed self-attention, turns a 224×224 image into 49 tokens, and a four-layer transformer decoder autoregressively predicts the plate string until [EOS], so both lines of a double-line plate are read as one spatial token layout rather than a flattened sequence. Trained on the corrected, extended mixed CCPD-derived test set, TransLPRNet reports 98.75% average accuracy under coarse bounding-box perturbations and 99.03% under fine four-vertex perturbations, beating Eulpr, PaddleOCRv3, Lpr-transformer, and TrOCR-without-pretraining on most subsets at 46 FPS with 5.94M parameters. PTN regresses the four corner coordinates, solves the 8-parameter perspective matrix, and is supervised by a frozen front-view classifier, improving coarse-perturbation accuracy from 98.60% to 98.91% and lifting the Challenge subset from 93.25% to 94.54%.

Load-bearing premise

The load-bearing premise is that the synthetic double-line plates used for training and testing faithfully represent real double-line plates in font, wear, distortion, and layout, and that the manual re-annotation of CCPD labels is accurate; if either fails, the reported recognition rates will not transfer to real plates.

Editorial extensions

If this is right

One model can recognize both single- and double-line Chinese plates without line segmentation or a two-stage pipeline; the decoder simply reads both lines as one token sequence.
PTN's weak-supervised corner regression is a stable alternative to STN-based perspective correction: under coarse localization perturbations, the paper reports average accuracy rising from 98.60% to 98.91%, with the hardest Challenge subset rising from 93.25% to 94.54%.
At 46 FPS and 5.94M parameters (batch size 1), the reported accuracy does not come at the cost of real-time deployment; the paper's comparison with TrOCR shows 12 FPS and 34.6M parameters.
Redundancy compression of the CCPD Base subset allows double-line plate images to be added without enlarging the dataset or sacrificing single-line information, so the same data budget supports both plate types.

Reading between the lines

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

A decisive test of the central transfer claim would be to evaluate the pipeline on a corpus of real (non-synthesized) double-line plates from truck or bus surveillance; if accuracy drops materially below the reported ~99% Base-d figure, the synthetic dataset's fidelity is the binding constraint.
The same weak-supervision recipe—a cheap binary 'is this a frontal view?' classifier driving a geometric rectifier—could be applied to ID cards, documents, or industrial markings where corner labels are scarce.
Since the model is a generic image-to-sequence network, fine-tuning it for other multi-line plate formats or for two-row document text is a natural extension; the bottleneck is likely data, not architecture.
The paper's error analysis points to two cheap improvements it does not test: a language or plate-format prior to disambiguate 8/B, 2/Z, 0/D, and 5/S confusions, and rebalancing the provincial distribution to reduce over-prediction of 'wan' from Anhui-dominant CCPD data.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's my take on TransLPRNet.

The core method is a straightforward recombination — MobileViTv3 encoder plus a small transformer decoder, with a Spatial-Transformer-style correction module — but the PTN design is the one genuinely interesting piece. Instead of regressing a perspective matrix directly, it regresses four corner points and uses a frozen frontal-view classifier as weak supervision. That's a clever end-run around the documented STN instability, and the ablation (Table 8) gives it modest, believable credit: +0.3% average on coarse localization. I'd call that a real contribution, even if not a scientific breakthrough.

The dataset work is also reasonable. The redundancy-compression trick — using an RPNet baseline to pick 'easy' CCPD images, then pasting synthetic double-line plates onto those backgrounds — is a sensible way to build a mixed dataset without bloating it, and the provincial-balance plots suggest care.

Now the soft spots, in proportion. The biggest one is external validity: the Base-d test set is entirely synthetic, generated by the same pipeline used to create the training double-line plates. So the 99.97% Base-d accuracy measures the model's ability to recognize the synthetic generator's output, not real double-line plates with real wear, font variation, and dirt. That's a load-bearing gap for the paper's main practical claim, and it should be stated as a limitation rather than buried.

Second, the numbers don't align. The arXiv abstract reports 99.34%/99.58% and 167 FPS; the paper's own abstract and tables report 98.75%/99.03% and 46 FPS. One of those is wrong. Also, Section 4.3.1 claims the model is trained from scratch with no external pre-trained data, but Section 3.1.1 says it uses ImageNet-1k pretrained weights for the backbone. That's a direct contradiction that needs fixing.

Third, no code, data, or corrected labels are released, so independent verification is impossible right now.

On balance, though, the engineering is coherent, the comparisons are run under one protocol, and the tables are internally consistent. The paper deserves a proper review — it just shouldn't be accepted as-is. A revised version that validates on real double-line plates, aligns the abstract, resolves the pretraining contradiction, and releases the synthetic pipeline would be a useful contribution.

Referee Report

4 major / 5 minor

Summary. The paper proposes TransLPRNet, a lightweight Chinese license plate recognition network combining a MobileViTv3-based visual encoder with a transformer text decoder, a Perspective Transformation Network (PTN) trained via weak supervision from a frontal-view classifier, and a dataset-construction pipeline that overlays program-generated double-line plates onto redundant CCPD images. The authors report averages of 98.75% and 99.03% recognition accuracy on a mixed single/double-line test set under coarse and fine localization perturbations, respectively, with 46 FPS and 5.94M parameters, outperforming LPRNet, Eulpr, PaddleOCRv3, TrOCR, and Lpr-transformer on most subsets.

Significance. If the results hold, the work is a useful engineering contribution: it demonstrates that a compact encoder-decoder can handle both single- and double-line Chinese plates and that viewpoint-classification supervision can train a perspective-correction module at low annotation cost. The body's tables are internally consistent, and the main comparisons are run under one protocol with reported speed and parameter counts. However, the central double-line claim is validated only on synthetic test images generated by the same pipeline used for training; the paper also contains a direct contradiction about whether TransLPRNet uses pre-trained weights; and the abstract in the metadata reports different headline numbers from the body. These issues prevent the evidence from supporting the paper's strongest claims as currently presented.

major comments (4)

[Section 3.3.2, Tables 6/7] The double-line recognition claim rests entirely on synthetic test images: the Base-d test set is built by pasting program-generated double-line plates onto redundant CCPD backgrounds, using the same generator that produced the double-line training images. The reported 99.97% and 99.99% Base-d accuracies are therefore within-distribution scores on synthetic plates, not evidence of generalization to real double-line plates, which differ in font rendering, embossing, wear, dirt, and layout. The paper should add a real double-line test set or an independent dataset, or clearly re-scope the claim away from 'strong practical applicability' in the abstract.
[Section 3.1.1 vs. Section 4.3.1] The paper contradicts itself on pre-training. Section 3.1.1 states that the visual encoder 'leveraged the pre-trained weights of this backbone on the ImageNet-1k dataset,' while Section 4.3.1 claims that TransLPRNet is 'trained entirely from scratch without using any external pre-trained data.' The comparison against TrOCR is explicitly framed as showing superiority without pre-training, so this contradiction is load-bearing for the fairness argument and must be resolved with a precise statement of which weights are initialized from ImageNet-1k and what is trained from scratch.
[Abstract (metadata vs. body)] The abstract in the submission metadata reports 99.34% average recognition accuracy under coarse localization, 99.58% under fine localization, and 167 FPS, whereas the body abstract and Tables 6/7 report 98.75% and 99.03% accuracy and 46 FPS. These are inconsistent headline results for the same claimed system, and the discrepancy is not explained anywhere in the manuscript. The authors must reconcile or clearly label which numbers correspond to which configuration.
[Section 3.3.1] The test labels are corrected using a model-based filtering step followed by manual re-annotation, and the corrected dataset is not released. Since all test accuracies depend on these modified labels, the paper should report the reliability of the correction process (e.g., inter-annotator agreement on a sample) and make the corrected labels available, otherwise independent verification of the headline accuracy is impossible.

minor comments (5)

[Section 3.2.2, Eq. (1)] The unknown parameter vector is ordered as [theta1, theta4, theta7, theta2, theta5, theta8, theta3, theta6], which does not match the row-by-row ordering implied by the coefficient matrix; please clarify or re-order the entries to avoid an apparent typographical inconsistency.
[Tables 4 and 5] The column header 'Parameter Name' is used for values such as Learning Rate, Optimizer, and Batch Size; rename it to 'Value' or 'Setting' for clarity.
[Tables 6 and 7] The 'Avg' column is not defined; please state whether it is a weighted average over subset sizes or an unweighted mean.
[Figure 18 and Section 3.3.2] The pie chart labels (e.g., 60k blue single-line, 10k green double-line, 10k yellow double-line) are difficult to reconcile with the text describing 25,000 double-line images added to the training set and 25,000 used in Base-d; clarify the exact dataset composition and the fate of the replaced single-line images.
[Throughout] There are numerous typographical issues, including 'Toaddress' in the abstract, 'Inversted' in Figure captions, inconsistent capitalization of 'transLPRNet' vs. 'TransLPRNet', and an unclear explanation in the Table 7 note about TrOCR's performance under fine localization; these should be corrected in a final revision.

Circularity Check

1 steps flagged · score 3.0 of 10

No formal circularity: single-line CCPD results are externally benchmarked, and no fitted parameter is renamed a prediction. One circularity-adjacent step remains: the synthetic Base-d test set is generated by the same Section 3.3.2 pipeline as the double-line training data, so the headline 99.97% Base-d accuracy measures intra-generator fit, not real double-line generalization.

other [Section 3.3.2 (Double/Single-Line Dataset Construction via Image Overlay); Tables 6-7; Abstract.]
"First, the program-generated double-line license plate images are blurred ... these synthesized double-line license plate images are overlaid onto redundant license plate samples via perspective transformation ... Among these, 25,000 double-line plates were added to the training set ... The remaining 25,000 double-line license plates were used as the Base-d subset within the test set to evaluate the model's recognition performance on double-line license plate images."

The double-line training and Base-d test plates are all rendered by the same Section 3.3.2 generator: program-made templates, blur, and perspective pasting onto RPNet-selected 'redundant' CCPD backgrounds. The Base-d test distribution is therefore the same distribution the model was trained on, so the headline double-line accuracies (99.97% coarse / 99.99% fine, Tables 6/7) certify intra-generator fit, not transfer to real double-line plates, of which the paper presents no evaluation; the abstract's 'strong practical applicability' overreaches this evidence.

full rationale

Formal derivation chain is self-contained. TransLPRNet (MobileViTv3 encoder + transformer decoder) is trained on the mixed CCPD/synthetic dataset; the PTN rectifier is trained under a frozen frontal-view classifier that itself was trained on labeled frontal/non-frontal plate images, so the PTN supervision signal is independent of the recognition target. Recognition is evaluated against external open-source baselines (LPRNet, Eulpr, PaddleOCRv3, TrOCR, Lpr-transformer) on identical test sets with identical Gaussian perturbations, so the single-line comparisons do not reduce to the paper's own outputs, and no fitted parameter is renamed as a prediction. There are no load-bearing author self-citations and no imported uniqueness theorems; architecture choices are cited from external works (MobileViTv3, TrOCR). The single circularity-adjacent issue is the double-line evaluation circularity in Section 3.3.2 described in the step above: because the Base-d test set comes from the same generator as the double-line training data, the double-line accuracy claims (99.97%/99.99%) validate fit to the generator, and the paper supplies no real double-line validation, which weakens the 'strong practical applicability' claim for its central dual-style contribution; this supports score 3 rather than 0. Flagged separately as correctness/consistency risks, not circular steps: (i) the arXiv abstract reports 99.34%/99.58% on CCPD, 98.70% on double-line, and 167 FPS, none matching Tables 6/7 (98.75%/99.03%, Base-d 99.97%/99.99%, 46 FPS); (ii) Section 3.1.1 states ImageNet-1k pre-trained encoder weights are leveraged, while Section 4.3.1 states TransLPRNet is 'trained entirely from scratch without using any external pre-trained data'; (iii) Base-s is enriched with RPNet-correctly-recognized 'redundant' samples, and the assertion that 'difficult-to-recognize samples share common characteristics across different models' is assumed, not demonstrated; (iv) no code, data, or corrected CCPD labels are released, so the corrected test set and synthetic pipeline cannot be independently verified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the synthetic double-line data being realistic, the corrected CCPD labels being accurate, and the weak classifier signal being sufficient for PTN training. Several hand-chosen simulation parameters (perturbation std, rectified size, redundancy count, decoding length) also shape the reported numbers.

free parameters (4)

Gaussian perturbation std = 4 pixels
Standard deviation of the random offset added to bounding-box and vertex coordinates to simulate detection errors (Section 4.1). The reported accuracies depend on this arbitrary simulation strength.
Target rectified image size = 94x24
Fixed output size for PTN rectification (Table 4). The perspective transformation maps plates to this canonical size, influencing recognition input.
Redundant-sample selection count = 50,000 (25k train, 25k test)
Number of CCPD images selected for synthetic double-line overlay (Section 3.3.2). This choice controls the balance of the mixed dataset.
Max decoding length N = 10
Maximum number of characters the decoder outputs (Section 3.1.2). Set according to plate type.

assumptions (4)

domain assumption Synthetic double-line plates are representative of real double-line plates
The double-line test set is composed of generated plates pasted onto real backgrounds (Section 3.3.2). If this does not hold, the reported double-line accuracy is not meaningful for real-world deployment.
domain assumption The manually corrected CCPD labels are accurate
The authors re-annotated 1,414 images flagged by an IOU-based filter (Section 3.3.1). The evaluation assumes these corrected labels are ground truth.
domain assumption Front-view classification provides sufficient supervision for perspective rectification
PTN is trained via a frozen binary front-view classifier (Section 3.2.3). The method assumes this weak signal is enough to learn correct corner coordinates.
standard math Standard deep learning assumptions (SGD/Adam, loss gradients, i.i.d. data) hold
The paper relies on standard training procedures without formal guarantees.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TransLPRNet: Lite Vision-Language Network for Single/Dual-line Chinese License Plate Recognition." pith.science (2026). https://pith.science/paper/GLLQVNWS

@misc{pith2026250717335,
  author       = {Pith},
  title        = {Pith review of: TransLPRNet: Lite Vision-Language Network for Single/Dual-line Chinese License Plate Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GLLQVNWS}},
  note         = {Machine review of arXiv:2507.17335}
}

read the original abstract

License plate recognition in open environments is widely applicable across various domains; however, the diversity of license plate types and imaging conditions presents significant challenges. To address the limitations encountered by CNN and CRNN-based approaches in license plate recognition, this paper proposes a unified solution that integrates a lightweight visual encoder with a text decoder, within a pre-training framework tailored for single and double-line Chinese license plates. To mitigate the scarcity of double-line license plate datasets, we constructed a single/double-line license plate dataset by synthesizing images, applying texture mapping onto real scenes, and blending them with authentic license plate images. Furthermore, to enhance the system's recognition accuracy, we introduce a perspective correction network (PTN) that employs license plate corner coordinate regression as an implicit variable, supervised by license plate view classification information. This network offers improved stability, interpretability, and low annotation costs. The proposed algorithm achieves an average recognition accuracy of 99.34% on the corrected CCPD test set under coarse localization disturbance. When evaluated under fine localization disturbance, the accuracy further improves to 99.58%. On the double-line license plate test set, it achieves an average recognition accuracy of 98.70%, with processing speeds reaching up to 167 frames per second, indicating strong practical applicability.

Figures

Figures reproduced from arXiv: 2507.17335 by the authors.

**Figure 1.** System solution diagram Input Image Linear +Softmax Multi-Head Cross Attention Add & Norm Masked Multi-Head Attention Add & Norm FFN Add & Norm 皖 J M H 7 8 K EOS Conv-3×3 Inversted Residual Block ×5 Decoder output Positional Encoding Decoder ×4 Local Representation Unit Inversted Residual Block Conv-1×1 MobileViT Block ×2 ×4 Inversted Residual Block MobileViT Block ×3 Patch Embedding MobileViT Block Linear Input Ima… view at source ↗

**Figure 2.** TransLPRNet network structure diagram Guangzhu Xu et al.: Preprint submitted to Elsevier Page 16 of 26 [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

**Figure 3.** TransLPRNet encoder network structure diagram Linear Decoder module Decoder module ×4 Add & Norm Add & Norm Add & Norm Masked Multi- Head Self- Attention Multi-Head Cross Attention FFN Positional Encoding argmax V K Q V K Q 皖 J M 7 8 K 皖 J M H 7 8 K Encoder input 128 49 SOS H EOS Output Embedding Generate the sequence step by step until the end-of-sentence marker (EOS) is output. E E ① ② ③ [PITH_FULL_IMAGE:figures/… view at source ↗

Figures from the paper (26 more)

**Figure 4.** Figure 4: TransLPRNet decoder network structure diagram Guangzhu Xu et al.: Preprint submitted to Elsevier Page 17 of 26 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

**Figure 5.** Figure 5: PTN correction network system block diagram 24×94×3 22×92×32 11×46×32 7×42×32 7×42×32 7×42×64 7×42×128 7×42×64 7×42×32 Conv MaxPool FCL 896 2084 32 （X1，Y1）（X2，Y2）（X3，Y3）（X4，Y4） [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

**Figure 6.** Figure 6: Vertex regression subnetwork (Um1，Vm1) (Um2，Vm2) (Um3，Vm3) (Um4，Vm4) (Xn1，Yn1) (Xn2，Yn2) (Xn3，Yn3) (Xn4，Yn4) 𝜃1 𝜃2 𝜃3 𝜃4 𝜃5 𝜃6 𝜃7 𝜃8 1 ( ) Xni Yni , ( ) Uni Vni , A [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

**Figure 7.** Figure 7: Mapping diagram between the LP four vertices and the four corner points of the input or output image Guangzhu Xu et al.: Preprint submitted to Elsevier Page 18 of 26 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

**Figure 8.** Figure 8: illustration of PTN supervision signals from the frontal license plate image classifier [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

**Figure 9.** Figure 9: Illustration of license plate images before label correction [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

**Figure 10.** Figure 10: Illustration of license plate images after label correction [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

**Figure 12.** Figure 12: Flowchart of composite training dataset construction Generated green doubleline license plate Green double-line license plate blurred Generated yellow doubleline license plate Yellow double-line license plate blurred [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]

**Figure 13.** Figure 13: Schematic diagram of double-line license plates and blurring [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]

**Figure 14.** Figure 14: Example of cropped images from the double-line license plate test set Guangzhu Xu et al.: Preprint submitted to Elsevier Page 20 of 26 [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]

**Figure 15.** Figure 15: Histogram of provincial distribution before dataset augmentation 皖苏浙沪豫粤京鄂鲁闽冀赣川湘渝晋津辽陕蒙黑甘云贵新桂吉青琼宁藏 [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]

**Figure 16.** Figure 16: Histogram of provincial distribution after dataset augmentation CCPD2019 Multi-type license plate training set Multi-type license plate test set [PITH_FULL_IMAGE:figures/full_fig_p021_16.png]

**Figure 17.** Figure 17: Example of overlaying generated double-line license plates onto CCPD dataset images Guangzhu Xu et al.: Preprint submitted to Elsevier Page 21 of 26 [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]

**Figure 18.** Figure 18: Pie chart of the distribution of the single/double-line license plate training set and single/double-line license plate test set (a) Bounding box localization (b) Bounding box localization disturbance (d) Four Vertex location disturbance (c) Four Vertex location [PIT…

**Figure 19.** Figure 19: Illustration of coarse and fine perturbations in license plate localization PTN  Classification Network Training 1 Lr=0.001 PTN Classification Network Training 2 Lr=0.001 Feedback   Train Freeze [PITH_FULL_IMAGE:figures/full_fig_p022_19.png]

**Figure 20.** Figure 20: Training workflow of the PTN Guangzhu Xu et al.: Preprint submitted to Elsevier Page 22 of 26 [PITH_FULL_IMAGE:figures/full_fig_p022_20.png]

**Figure 21.** Figure 21: Example of STN before and after affine transformation correction [PITH_FULL_IMAGE:figures/full_fig_p023_21.png]

**Figure 22.** Figure 22: Example images before and after STN perspective transformation correction (a) Before PTN correction (b) After PTN correction (a) Before correction (b) After correction [PITH_FULL_IMAGE:figures/full_fig_p023_22.png]

**Figure 23.** Figure 23: Comparison of images before and after PTN correction (a) Before PTN correction (b) After PTN correction (a) Before correction (b) After correction [PITH_FULL_IMAGE:figures/full_fig_p023_23.png]

**Figure 24.** Figure 24: Experimental results of PTN integrated with TransLPRNet Guangzhu Xu et al.: Preprint submitted to Elsevier Page 23 of 26 [PITH_FULL_IMAGE:figures/full_fig_p023_24.png]

**Figure 25.** Figure 25: Data augmentation visualization Label : 皖AEE925 Predict: 皖AEE925 Label : 鄂A9XT19 Predict: 鄂A9XT19 Label : 京LD7865 Predict: 京LD7865 Label : 皖AS5596 Predict: 皖AS5596 Label : 皖AD4830 Predict: 皖AD4830 Label : 皖AC808P Predict: 皖AC808P Label : 皖A68G72 Predict: 皖A68G72 Label…

**Figure 26.** Figure 26: Visualization of single-line license plate image recognition results Label : 蒙XR6CN3 Predict: 蒙XR6CN3 Label : 藏CZ952T Predict: 藏CZ952T Label : 川B20H6B Predict: 川B20H6B Label : 沪MAT98U Predict: 沪MAT98U Label : 川32S3K8X Predict: 川32S3K8X Label : 陕08KPKA8 Predict: 陕08KPK…

**Figure 27.** Figure 27: Visualization of double-line license plate image recognition results Guangzhu Xu et al.: Preprint submitted to Elsevier Page 24 of 26 [PITH_FULL_IMAGE:figures/full_fig_p024_27.png]

**Figure 28.** Figure 28: Comparison of recognition results with and without using PTN Label : 皖AD515W Predict: 皖A0515W Label : 皖AZ8355 Predict: 皖A28355 Label : 皖A59H88 Predict: 皖A59HB8 Label : 皖AT319S Predict: 皖AT3195 Label : 皖AD601Z Predict: 皖A0601Z Label : 皖A592T6 Predict: 皖A59ZT6 Label : 皖…

**Figure 29.** Figure 29: Illustration of license plate character confusion and recognition errors 皖A5K223 皖K41273 皖A4F605 皖AH1305 皖A8525H 皖A85231 皖A79Z76 皖AZ0276 Recognition result Before correction After correction Recognition result Label 皖B21273 皖AHF305 皖AB523H 皖A7P276 Label 皖B21273 皖AHF30…

**Figure 30.** Figure 30: Illustration of license plate character confusion and recognition errors Guangzhu Xu et al.: Preprint submitted to Elsevier Page 25 of 26 [PITH_FULL_IMAGE:figures/full_fig_p025_30.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages

[1]

License plate recognition system based on improved yolov5 and gru

Shi, H., Zhao, D., 2023. License plate recognition system based on improved yolov5 and gru. Ieee Access 11, 10429–10439

work page 2023
[2]

Improving robustness of license plates automatic recognition in natural scenes

Fan, X., Zhao, W., 2022. Improving robustness of license plates automatic recognition in natural scenes. IEEE Transactions on Intelligent Transportation Systems 23, 18845–18854

work page 2022
[3]

Vehicle license plate recognition method based on deep convolution network in complex road scene

Liu, Z., Cai, Y., Chen, L., Wang, H., He, Y., 2019. Vehicle license plate recognition method based on deep convolution network in complex road scene. Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering 233, 2284–2292

work page 2019
[4]

A new approach for character recognition of multi-style vehicle license plates

Huang, Q., Cai, Z., Lan, T., 2020. A new approach for character recognition of multi-style vehicle license plates. IEEE Transactions on multimedia 23, 3768–3777

work page 2020
[5]

Robust automatic recognition of chinese license plates in natural scenes

He, M.X., Hao, P., 2020. Robust automatic recognition of chinese license plates in natural scenes. Ieee Access 8, 173804–173814

work page 2020
[6]

License plate detection and recognition based on yolov3 and ilprnet

Zou, Y., Zhang, Y., Yan, J., Jiang, X., Huang, T., Fan, H., Cui, Z., 2022. License plate detection and recognition based on yolov3 and ilprnet. Signal, image and video processing 16, 473–480

work page 2022
[7]

A robust license plate recognition model based on bi-lstm

Zou, Y., Zhang, Y., Yan, J., Jiang, X., Huang, T., Fan, H., Cui, Z., 2020. A robust license plate recognition model based on bi-lstm. IEEE Access 8, 211630–211641

work page 2020
[8]

Lprnet: License plate recognition via deep neural networks

Zherzdev, S., Gruzdev, A., . Lprnet: License plate recognition via deep neural networks. arxiv 2018. arXiv preprint arXiv:1806.10447

arXiv 2018

Show all 29 references

[9]

Recognitionofvehiclelicenseplatesinhighwaysceneswithdeepfusionnetwork and connectionist temporal classification

Hua,L.,Ma,X.,Zhao,C.,Zhang,B.,Su,Z.,Wu,Y.,2024. Recognitionofvehiclelicenseplatesinhighwaysceneswithdeepfusionnetwork and connectionist temporal classification. IET Image Processing 18, 4066–4080

2024
[10]

License plate recognition system using yolov5 and cnn, in: 2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS), IEEE

Raj, S., Gupta, Y., Malhotra, R., 2022. License plate recognition system using yolov5 and cnn, in: 2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS), IEEE. pp. 372–377

2022
[11]

Automatic number plate recognition (anpr) with yolov3-cnn

Adak, R., Kumbhar, A., Pathare, R., Gowda, S., 2022. Automatic number plate recognition (anpr) with yolov3-cnn. arXiv preprint arXiv:2211.05229

2022 arXiv
[12]

Spatial transformer networks

Jaderberg, M., Simonyan, K., Zisserman, A., et al., 2015. Spatial transformer networks. Advances in neural information processing systems 28

2015
[13]

Collaborativelicenseplaterecognitionviaassociationenhancementnetwork with auxiliary learning and a unified benchmark

Deng,Y.,Wang,G.,Li,C.,Wang,W.,Zhang,C.,Tang,J.,2024. Collaborativelicenseplaterecognitionviaassociationenhancementnetwork with auxiliary learning and a unified benchmark. IEEE Transactions on Multimedia

2024
[14]

A deep learning-based framework for vehicle license plate detection

Yang, D., Yang, L., 2024. A deep learning-based framework for vehicle license plate detection. International Journal of Advanced Computer Science & Applications 15

2024
[15]

Trocr:Transformer-basedopticalcharacterrecognition with pre-trained models, in: Proceedings of the AAAI conference on artificial intelligence, pp

Li,M.,Lv,T.,Chen,J.,Cui,L.,Lu,Y.,Florencio,D.,Zhang,C.,Li,Z.,Wei,F.,2023. Trocr:Transformer-basedopticalcharacterrecognition with pre-trained models, in: Proceedings of the AAAI conference on artificial intelligence, pp. 13094–13102

2023
[16]

Mobilevitv3: Mobile-friendly vision transformer with simple and effective fusion of local, global and input features

Wadekar, S.N., Chaurasia, A., 2022. Mobilevitv3: Mobile-friendly vision transformer with simple and effective fusion of local, global and input features. arXiv preprint arXiv:2209.15159

2022 arXiv
[17]

Towards end-to-end license plate detection and recognition: A large dataset and baseline, in: Proceedings of the European conference on computer vision (ECCV), pp

Xu, Z., Yang, W., Meng, A., Lu, N., Huang, H., Ying, C., Huang, L., 2018. Towards end-to-end license plate detection and recognition: A large dataset and baseline, in: Proceedings of the European conference on computer vision (ECCV), pp. 255–271

2018
[18]

Qin,S.,Liu,S.,2020.Efficientandunifiedlicenseplaterecognitionvialightweightdeepneuralnetwork.IETImageProcessing14,4102–4109

2020
[19]

Pp-ocr: A practical ultra lightweight ocr system

Du, Y., Li, C., Guo, R., Yin, X., Liu, W., Zhou, J., Bai, Y., Yu, Z., Yang, Y., Dang, Q., et al., 2020. Pp-ocr: A practical ultra lightweight ocr system. arXiv preprint arXiv:2009.09941

2020 arXiv
[20]

Pp-ocrv3:Moreattemptsfortheimprovement of ultra lightweight ocr system

Li,C.,Liu,W.,Guo,R.,Yin,X.,Jiang,K.,Du,Y.,Du,Y.,Zhu,L.,Lai,B.,Hu,X.,etal.,2022. Pp-ocrv3:Moreattemptsfortheimprovement of ultra lightweight ocr system. arXiv preprint arXiv:2206.03001

2022 arXiv
[21]

Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer

Mehta, S., Rastegari, M., 2021. Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178

2021 arXiv
[22]

Integrating yolov8 and cspbottleneck based cnn for enhanced license plate character recognition

Khokhar, S., Kedia, D., 2024. Integrating yolov8 and cspbottleneck based cnn for enhanced license plate character recognition. Journal of Real-Time Image Processing 21, 168

2024
[23]

Robust license plate detection and recognition with automatic rectification

Xiao, D., Zhang, L., Li, J., Li, J., 2021. Robust license plate detection and recognition with automatic rectification. Journal of Electronic Imaging 30, 013002–013002

2021
[24]

Bakshi, A., Gulhane, S., Sawant, T., Sambhe, V., Udmale, S.S., 2023. Alpr-an intelligent approach towards detection and recognition of licenseplatesinuncontrolledenvironments,in:InternationalConferenceonDistributedComputingandIntelligentTechnology,Springer.pp. 253–269

2023
[25]

A robust attentional framework for license plate recognition in the wild

Zhang, L., Wang, P., Li, H., Li, Z., Shen, C., Zhang, Y., 2020. A robust attentional framework for license plate recognition in the wild. IEEE Transactions on Intelligent Transportation Systems 22, 6967–6976

2020
[26]

Wang,Q.,Lu,X.,Zhang,C.,Yuan,Y.,Li,X.,2022.Lsv-lp:Large-scalevideo-basedlicenseplatedetectionandrecognition.IEEETransactions on Pattern Analysis and Machine Intelligence 45, 752–767

2022
[27]

Searching for mobilenetv3, in: Proceedings of the IEEE/CVF international conference on computer vision, pp

Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al., 2019. Searching for mobilenetv3, in: Proceedings of the IEEE/CVF international conference on computer vision, pp. 1314–1324

2019
[28]

Flexiblecameracalibrationbyviewingaplanefromunknownorientations,in:Proceedingsoftheseventhieeeinternational conference on computer vision, Ieee

Zhang,Z.,1999. Flexiblecameracalibrationbyviewingaplanefromunknownorientations,in:Proceedingsoftheseventhieeeinternational conference on computer vision, Ieee. pp. 666–673

1999
[29]

chinese-lpr-transformer

sosopop, 2024. chinese-lpr-transformer. GitHub. URL:https://github.com/sosopop/chinese-lpr-transformer. Guangzhu Xu et al.:Preprint submitted to ElsevierPage 26 of 26

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

License plate recognition system based on improved yolov5 and gru

Shi, H., Zhao, D., 2023. License plate recognition system based on improved yolov5 and gru. Ieee Access 11, 10429–10439

work page 2023

[2] [2]

Improving robustness of license plates automatic recognition in natural scenes

Fan, X., Zhao, W., 2022. Improving robustness of license plates automatic recognition in natural scenes. IEEE Transactions on Intelligent Transportation Systems 23, 18845–18854

work page 2022

[3] [3]

Vehicle license plate recognition method based on deep convolution network in complex road scene

Liu, Z., Cai, Y., Chen, L., Wang, H., He, Y., 2019. Vehicle license plate recognition method based on deep convolution network in complex road scene. Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering 233, 2284–2292

work page 2019

[4] [4]

A new approach for character recognition of multi-style vehicle license plates

Huang, Q., Cai, Z., Lan, T., 2020. A new approach for character recognition of multi-style vehicle license plates. IEEE Transactions on multimedia 23, 3768–3777

work page 2020

[5] [5]

Robust automatic recognition of chinese license plates in natural scenes

He, M.X., Hao, P., 2020. Robust automatic recognition of chinese license plates in natural scenes. Ieee Access 8, 173804–173814

work page 2020

[6] [6]

License plate detection and recognition based on yolov3 and ilprnet

Zou, Y., Zhang, Y., Yan, J., Jiang, X., Huang, T., Fan, H., Cui, Z., 2022. License plate detection and recognition based on yolov3 and ilprnet. Signal, image and video processing 16, 473–480

work page 2022

[7] [7]

A robust license plate recognition model based on bi-lstm

Zou, Y., Zhang, Y., Yan, J., Jiang, X., Huang, T., Fan, H., Cui, Z., 2020. A robust license plate recognition model based on bi-lstm. IEEE Access 8, 211630–211641

work page 2020

[8] [8]

Lprnet: License plate recognition via deep neural networks

Zherzdev, S., Gruzdev, A., . Lprnet: License plate recognition via deep neural networks. arxiv 2018. arXiv preprint arXiv:1806.10447

arXiv 2018