Pith. sign in

REVIEW 3 major objections 7 minor 34 references

LPTR-AFLNet: Lightweight Integrated Chinese License Plate Rectification and Recognition Network

T0 review · 3 major / 7 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 2.7 MB end-to-end network unifies perspective rectification and recognition of single- and double-line Chinese license plates by letting the recognizer's own output weakly supervise the geometry correction.

desk verdict A competent lightweight LPR pipeline with a clever double-line rectification trick, but the headline double-line accuracy is measured only on synthetic data from the same generator, so the central claim needs independent validation before it can be trusted. read the letter →

arxiv 2507.16362 v3 pith:QTYLBFXU submitted 2025-07-22 cs.CV

Guangzhu Xu , Pengcheng Zuo , Zhi Ke , Bangjun Lei This is my paper · ORCID
classification cs.CV
keywords ChineseLicensePlateRecognitionPerspectiveRectificationLightweightNetworkAttentionModuleFocalCTCLossDouble-lineWeakSupervisionLPRNet
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

The paper is trying to establish that a single lightweight network can do both jobs that normally require separate, heavier modules: undo perspective distortion and recognize Chinese plates in single- and double-line formats. Its route is to let the recognizer itself supervise the geometry correction, so the rectification module learns to make plates readable rather than to match annotated vertices. On the CCPD benchmark it reports higher accuracy than LPRNet on every subset, including 98.87% average, 99.15% on Rotate, 99.19% on Tilt, and 90.39% on Challenge, plus 99.37% on a 20,000-image synthetic double-line test set, at 107 FPS on a TITAN X and 2.7 MB model size. If true, this matters because edge devices cannot run the heavy CRNN and transformer models that currently top benchmarks, and double-line plates, common in practice, have almost no public training data. The paper also contributes a 200,000-image synthetic double-line dataset and shows that perspective rectification, per-channel attention, and Focal CTC each add accuracy on top of LPRNet.

What carries the argument

The load-bearing object is the PTR module, a perspective transformation rectifier composed of a vertex-offset estimation subnet, a matrix solver, a grid generator, and a bilinear sampler. Instead of regressing perspective matrix parameters directly, which the paper argues makes STN training diverge, the subnet regresses vertex offsets and the matrix solver obtains the transform by solving the classic eight-point correspondence system. Its training signal is the recognition network's Focal CTC loss, making the whole loop end-to-end and weakly supervised. On the recognition side, LP-CA (lightweight per-channel attention) average-pools each of the 73 character-lexicon channels along height, applies a per-channel 1×3 convolution, and multiplies the sigmoid output back onto the features, suppressing the inter-character spacing confusions shown in the error analysis. Focal CTC replaces vanilla CTC by weighting the CTC loss with $\alpha(1-p)^\gamma$, where $p$ is the probability of the full correct plate. The double-line extension regresses twelve offsets that define six vertices, rectifies upper and lower regions with two matrices, and horizontally concatenates them.

What would settle it

Collect a few thousand real double-line Chinese plates with ground-truth text from multiple provinces, run the trained model without fine-tuning, and compare recognition accuracy against the 99.37% reported on the synthetic test set. A large drop, especially on characters outside the synthetic template set or on plates with real blur and weathering, would show that the rectifier and recognizer learned synthetic statistics rather than true geometry, whereas single-line CCPD accuracy staying high would confirm that the failure is specific to the double-line transfer assumption.

Watch

Extended reading notes

Core claim

The central claim is that perspective rectification of license plates becomes stable and annotation-free if the network regresses vertex offsets rather than perspective matrix entries. The PTR module estimates the eight offsets of the plate's four vertices relative to the input corners, solves the standard linear system for the perspective matrix, and samples the corrected image; training is driven entirely by the recognition loss, so the recognizer teaches the rectifier. For double-line plates the same idea is extended by estimating six shared vertices that split the plate into upper and lower regions, rectifying each region separately, and concatenating the two horizontally to form a single line. The recognition network starts from LPRNet and adds two cheap modifications: LP-CA, a per-channel attention that sharpens the 73 lexicon pages of high-level features, and Focal CTC, which reweights hard-to-recognize samples to counter the CCPD dataset's overwhelming 'WAN' majority. The paper claims this combination outperforms LPRNet on all CCPD subsets and reaches 99.37% on the author-built synthetic double-line test set.

Load-bearing premise

The double-line results depend on the assumption that the synthetic double-line dataset, made by blurring, warping, and compositing hand-crafted templates onto CCPD backgrounds, captures the distribution of real double-line plates closely enough that accuracy measured on it (99.37%) predicts accuracy in deployment.

Editorial extensions

If this is right

  • If the reported CCPD numbers hold, a 2.7 MB model can match or beat LPRNet on every CCPD subset while running at 107 FPS, so real-time edge deployment does not require giving up perspective robustness.
  • Double-line plates become trainable without a dedicated real-world corpus: the synthetic dataset plus recognition-supervised rectification is enough to reach 99.37% on the author's test split, and the unified pipeline removes the need for a separate single/double-line classifier with a tilt-sensitive slicing parameter.
  • Weak supervision from recognition means the rectifier can be trained on images with only bounding-box-level localization, avoiding costly vertex annotation for single-line plates.
  • The Focal CTC modification alone is reported to raise accuracy on plates with rare Chinese characters by 17%, implying that data imbalance, not just model capacity, is a major source of province-character errors in LPRNet.
  • The error statistics table shows the combined model cuts total CCPD misidentifications from 15,660 to 4,034, with the largest absolute reductions in missing characters and Chinese-character errors.

Reading between the lines

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

  • Editorial inference: the same vertex-offset-plus-recognition-loss recipe could transfer to other fixed-layout recognition tasks, such as ID cards, invoices, or shipping labels, where the recognizer's loss can supervise geometry correction without vertex annotations.
  • Editorial inference: because the CCPD benchmark is dominated by 'WAN' plates, the numeric gains on CCPD may understate the value of Focal CTC on a geographically balanced set; testing on a multi-province corpus would isolate that effect more cleanly.
  • Editorial inference: if the synthetic double-line data transfers as assumed, adding a small set of real double-line plates for final fine-tuning is a natural next step that could close any residual domain gap without retraining the whole network.
  • Editorial inference: the LP-CA operator is a generic per-channel temporal attention that could be dropped into other CTC-based OCR models at small parameter cost, so its benefit is testable independently of license plates.
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.

Referee Report

3 major / 7 minor

Summary. This paper presents LPTR-AFLNet, a lightweight end-to-end network that jointly rectifies and recognizes Chinese license plates for both single- and double-line formats. A perspective transformation rectification (PTR) module estimates per-vertex offsets and solves for a homography instead of regressing the perspective matrix directly, trained with weak supervision from the recognition head. For double-line plates the network estimates six vertices, rectifies the two lines separately, and concatenates them into one line before recognition. The recognition backbone is LPRNet augmented with a lightweight per-channel attention (LP-CA) module and a Focal CTC loss. The authors construct a 200,000-image synthetic double-line dataset from hand-crafted templates, blur, perspective warping, and CCPD background compositing, and they evaluate on CCPD, LSV, and their synthetic double-line test set, reporting accuracy above 99% on several CCPD subsets and 99.37% on the synthetic double-line set at 107 FPS on a TITAN X.

Significance. Assuming the results hold, the paper makes a useful engineering contribution: the PTR formulation avoids known instabilities of STN-based perspective regression, the double-line concatenation design is simple and elegant, and the staged weak-supervision training is clearly described. The large synthetic dataset and the detailed qualitative error analysis (character omission/insertion, Chinese-character imbalance) are valuable parts of the paper. The single-line CCPD results are plausible and consistent with the LPRNet baseline. However, the significance is bounded by three limitations: no code or data are released, all tables are single-run point estimates, and the double-line claim rests entirely on a synthetic test set generated by the same pipeline as the training data. The paper is therefore a solid candidate after major revisions rather than an accept in its current form.

major comments (3)
  1. [§3.4, Table 2] The headline double-line result (Double=99.37 in Table 2) is computed on the held-out 10% of the 200,000-image synthetic dataset built in §3.4 with hand-crafted templates, random blur, perspective warping from CCPD vertex coordinates, and compositing onto CCPD backgrounds. Since the same generator and the same CCPD-derived geometry are used for training and test, the number measures fit to the synthetic distribution rather than performance on real double-line plates. Real double-line plates (mostly trucks and trailers) differ in fonts, plate proportions, mounting angles, materials, and background/lighting; the paper's own error analysis in §4.4 shows sensitivity to non-Anhui characters, yellow plates, and blur. Without an independent real-world double-line evaluation, or an explicit restriction of the claim to the synthetic distribution, the central claim that LPTR-AFLNet 'excels' at double-line recognition is unvalidated.
  2. [Tables 2, 3, 5, 6] All accuracy numbers are single-run point estimates with no error bars, repeated-seed statistics, or significance tests. In the CCPD Base ablation (Table 5), PTR+LPRNet+LP-CA (99.39) and PTR+LPRNet+Focal CTC (99.35) differ by 0.04 percentage points, and the full model is only 0.07 points above the LP-CA-only variant; with roughly 100,000 test images and highly correlated CTC predictions, these differences are within plausible run-to-run noise. The claim that LP-CA and Focal CTC each contribute positively, and comparably, is therefore not established at the reported precision. The authors should report mean plus/minus standard deviation over multiple seeds, or a significance test, for the key ablations.
  3. [§4.3.2, Table 2] The comparison protocol in Table 2 is not apples-to-apples. The table is titled as a CCPD comparison but contains a 'Double' column that is not part of CCPD; the text does not state which double-line test set was used or whether comparison systems (e.g., Eulpr with 98.73) were evaluated on the same synthetic split. Moreover, models with unavailable code are said to recognize vertex-localized crops, while the proposed method is evaluated with random perturbation on the entire sample set, so differences between the upper and lower parts of the table mix algorithmic performance with input-resolution and localization-protocol differences. The authors need to specify the exact protocol per model, including the double-line test split, and define the 'AP' column.
minor comments (7)
  1. [§4.3.2, Table 4] The abstract's claim of 'less than 10 milliseconds' on 'lower-mid-range GPUs' is not supported by Table 4: the only GPU row is a TITAN X (107 FPS, about 9.3 ms), while the 79/50/33 FPS rows are CPU evaluations. Either report a lower-mid-range GPU result or revise the wording.
  2. [§3.3.2, Eq. (7)] The definition of p in Eq. (7) as the product of per-character greedy CTC probabilities is not the CTC posterior for the label sequence; please clarify whether this is an intentional sample-difficulty heuristic and, if so, state that the focal weighting is applied to the whole CTC loss rather than derived from per-path reweighting.
  3. [Table 2] The 'AP' column is never defined; the values appear to be a weighted average over the CCPD test subsets, but the paper should state the formula explicitly.
  4. [Figures 3 and 7] The label 'ΔY14' in Figures 3 and 7 appears to be a typo for 'ΔY4'; please correct.
  5. [§4.3.3] The text says the LSV ablation is run on the 'Move and Static' subsets, while Table 6 reports an Accuracy_7c of 41.26 for LPRNet, matching the 'Move vs Static' row in Table 3; please align the terminology.
  6. [Table 1] The hyperparameter table lists Adam as the optimizer alongside 'Momentum 0.9'; Adam does not take a momentum argument, so please specify whether SGD with momentum is used for any training stage or remove the entry.
  7. [General] No code or data availability statement is provided; given that the synthetic double-line dataset is central to the paper, releasing the generation code would materially improve reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-circularity in the core PTR/AFLNet pipeline; single-line claims rest on external CCPD/LSV benchmarks. The double-line accuracy is self-referential because the test set is generated by the same synthetic pipeline used for training.

  1. other [Section 3.4 (Double-line License Plate Dataset Construction) and Table 2 (Double column)]
    "First, to ensure the diversity of generated characters, we utilized meticulously designed hand-crafted templates to generate a large number of double-line license plate images ... Finally, to simulate multi-angle perspectives and complex lighting conditions in open environments, we referenced the license plate vertex coordinates provided in the CCPD dataset to apply perspective transformations to the generated double-line license plate images, overlaying the transformed images into corresponding CCPD sample images. ..."

    The 99.37% Double accuracy in Table 2 is computed on the held-out slice of the same hand-crafted-template/blur/perspective-warp/composite pipeline that produced the training images. Thus the headline double-line result measures consistency with the authors' own synthetic generator, not generalization to an independent real double-line distribution. The conclusion that LPTR-AFLNet 'excels in handling single/double-line license plate images' draws its double-line support from a benchmark that shares its construction pipeline with the training data, making that portion of the evaluation self-referential. This does not affect the external CCPD/LSV single-line comparisons, so the circularity is localized and partial.

full rationale

Walking the claimed derivation chain, the core PTR module computes a perspective matrix by solving Equation 1 from estimated vertex offsets, and the recognition network is trained with Focal CTC loss against the CCPD/LSV labels; no step reduces an output to its own input by construction. The weak-supervision loop, where recognition loss backpropagates through the rectification module, is a standard end-to-end training loop rather than a circular derivation. The single-line performance claims are grounded in external benchmarks (CCPD subsets and LSV), and the ablations isolate the contributions of PTR, LP-CA, and Focal CTC. The only self-referential element is the double-line evaluation: both training and test double-line images are generated by the same hand-crafted-template and compositing pipeline, so the Double 99.37% is partly self-confirming with respect to that synthetic distribution. There is also a self-citation ([32]) to the authors' prior YOLO-based detection work, but it is illustrative in the perturbation discussion and not load-bearing. Accordingly, the paper has no significant derivation circularity; the localized self-referential evaluation warrants a score of 2 rather than 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The paper's central single-line claim rests on standard homography geometry plus an empirical training pipeline; no physical constants or new forces are involved. The main ledger items are hyperparameters chosen by hand and the domain assumption that synthetic double-line data represents reality. The only invented entity is the synthetic double-line dataset, which lacks independent validation.

free parameters (4)
  • Focal CTC balance factor alpha = 0.5
    Chosen by hand in Section 3.3.2; the reported gains on rare Chinese characters can depend on this value.
  • Focal CTC adjustment factor gamma = 2.0
    Chosen by hand in Section 3.3.2 following standard Focal Loss defaults; sensitivity of the 17% rare-character improvement is not analyzed.
  • Gaussian perturbation std for localization simulation = 4 pixels
    Set in Section 4.1 to simulate detector errors; this choice directly controls the difficulty of the rectification task and hence measured accuracy.
  • Label screening IoU threshold = 0.6
    Used in Section 3.4 to decide which CCPD labels are erroneous; changing it alters the corrected dataset and the resulting double-line composition.
assumptions (5)
  • domain assumption The synthetic double-line dataset distribution approximates real-world double-line plates.
    Section 3.4 builds the dataset from templates plus CCPD backgrounds; the test split inherits this assumption, so the 99.37 Double accuracy is only meaningful if the synthetic distribution generalizes.
  • standard math A planar license plate's perspective distortion is exactly modeled by an 8-parameter homography with the ninth parameter fixed to 1.
    The transformation matrix solver in Section 3.1.2 uses Formula 1, which assumes planar homography; this is standard for rigid plates but excludes non-planar bending.
  • domain assumption Recognition output provides a sufficiently strong weak supervisory signal to learn vertex offsets.
    Training stage 2 in Section 4.2 freezes the recognizer and trains PTR only from recognition loss; if this signal is too noisy or ambiguous, the rectification module may converge to a trivial or wrong transform.
  • domain assumption CCPD labels, after IoU-based screening and manual correction, are accurate enough for training and evaluation.
    Section 3.4 identifies 1,414 mislabeled images and corrects them; the remainder of CCPD is treated as ground truth, including the test subsets used for the headline accuracy numbers.
  • ad hoc to paper The greedy CTC probability product in Focal CTC loss is a valid proxy for sample difficulty.
    Equation 7 defines p as the product of probabilities from the greedy decode; this is a design choice specific to this paper and not derived from the Focal Loss framework.
invented entities (1)
  • Synthetic double-line Chinese license plate dataset (200,000 images)
    purpose: Provides training and test data for double-line plates, which lack large public real-world datasets.
    Constructed via hand-crafted templates overlaid on CCPD backgrounds (Section 3.4); no external real-world double-line benchmark validates that it captures real plate variability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LPTR-AFLNet: Lightweight Integrated Chinese License Plate Rectification and Recognition Network." pith.science (2026). https://pith.science/paper/QTYLBFXU

@misc{pith2026250716362,
  author       = {Pith},
  title        = {Pith review of: LPTR-AFLNet: Lightweight Integrated Chinese License Plate Rectification and Recognition Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QTYLBFXU}},
  note         = {Machine review of arXiv:2507.16362}
}
read the original abstract

Chinese License Plate Recognition (CLPR) faces numerous challenges in unconstrained and complex environments, particularly due to perspective distortions caused by various shooting angles and the correction of single-line and double-line license plates. Considering the limited computational resources of edge devices, developing a low-complexity, end-to-end integrated network for both correction and recognition is essential for achieving real-time and efficient deployment. In this work, we propose a lightweight, unified network named LPTR-AFLNet for correcting and recognizing Chinese license plates, which combines a perspective transformation correction module (PTR) with an optimized license plate recognition network, AFLNet. The network leverages the recognition output as a weak supervisory signal to effectively guide the correction process, ensuring accurate perspective distortion correction. To enhance recognition accuracy, we introduce several improvements to LPRNet, including an improved attention module to reduce confusion among similar characters and the use of Focal Loss to address class imbalance during training. Experimental results demonstrate the exceptional performance of LPTR-AFLNet in rectifying perspective distortion and recognizing double-line license plate images, maintaining high recognition accuracy across various challenging scenarios. Moreover, on lower-mid-range GPUs platform, the method runs in less than 10 milliseconds, indicating its practical efficiency and broad applicability.

Figures

Figures reproduced from arXiv: 2507.16362 by the authors.

Figure 1
Figure 1. Overall System Block Diagram 3. Method While lightweight license plate recognition models, exemplified by LPRNet, have achieved significant progress, challenges remain in effectively handling license plates with perspective distortion, accommodating diverse plate formats, and mitigating feature overlap and confusion during recognition. To address these issues, we propose a novel, lightweight, integrated rectificatio… view at source ↗
Figure 2
Figure 2. Network structure of perspective distortion rectification module for single-line license plate ΔX1 ΔY1 ΔX2 ΔY2 ΔY3 ΔX3 ΔX4 ΔY14 ΔX1 ΔY1 ΔX2 ΔY2 ΔY3 ΔX3 ΔX4 ΔY14 94 3 36 7 7 29 8 2 2 88 10 40 2 2 5 10 20 1000 32 8 MaxPool 8 5 5 44 14 10 MaxPool . . . ΔX1 ΔY1 ΔX1 ΔY1 ΔX4 ΔY4 ΔX4 ΔY4 . . . ΔX1 ΔY1 ΔX4 ΔY4 94 3 36 7 7 29 8 2 2 88 10 40 2 2 5 10 20 1000 32 8 MaxPool 8 5 5 44 14 10 MaxPool . . . ΔX1 ΔY1 ΔX4 ΔY4 ΔX1 ΔY1 ΔX… view at source ↗
Figure 3
Figure 3. Structure of license plate vertex offset estimation subnetwork 3.1.2. Transformation Matrix Solver Submodule This submodule calculates the four-vertex coordinates of the license plate region. It achieves this by adding the vertex offset values regressed by the vertex offset estimation submodule to the normalized coordinates of the four corner points of the license plate image. Specially,(𝑋1 , 𝑌1 ) = (0, 0) + (Δ𝑋1 , … view at source ↗
Figures from the paper (30 more)
Figure 4
Figure 4. Figure 4: Specific Flowchart of the Transformation Matrix Solver Module x1 x x2 y1 y y2 A3’ A1’ A4’ R2 A2’ R1 A’ A A’ Mθ Sampling Interpolation x1 x x2 y1 y y2 A3’ A1’ A4’ R2 A2’ R1 A’ A A’ Mθ Sampling Interpolation [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Grid Generation and Sampling Diagram ⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣ 𝑈1 𝑉1 1 0 0 0 −𝑋1𝑈1 −𝑋1𝑉1 0 0 0 𝑈1 𝑉1 1 −𝑌1𝑈1 −𝑌1𝑉1 𝑈2 𝑉2 1 0 0 0 −𝑋2𝑈2 −𝑋2𝑉2 0 0 0 𝑈2 𝑉2 1 −𝑌2𝑈2 −𝑌2𝑉2 𝑈3 𝑉3 1 0 0 0 −𝑋3𝑈3 −𝑋3𝑉3 0 0 0 𝑈3 𝑉3 1 −𝑌3𝑈3 −𝑌3𝑉3 𝑈4 𝑉4 1 0 0 0 −𝑋4𝑈4 −𝑋4𝑉4 0 0 0 𝑈4 𝑉4 1 −𝑌4𝑈4 −𝑌4𝑉4 ⎤ ⎥ …
Figure 6
Figure 6. Figure 6: Network structure of perspective distortion rectification module for double-line license plate 𝑓(𝑥, 𝑦1 ) ≈ 𝑥2 − 𝑥 𝑥2 − 𝑥1 𝑓(𝐴 ′ 1 ) + 𝑥 − 𝑥1 𝑥2 − 𝑥1 𝑓(𝐴 ′ 2 ) (2) 𝑓(𝑥, 𝑦2 ) ≈ 𝑥2 − 𝑥 𝑥2 − 𝑥1 𝑓(𝐴 ′ 3 ) + 𝑥 − 𝑥1 𝑥2 − 𝑥1 𝑓(𝐴 ′ 4 ) (3) 𝑓(𝑥, 𝑦) ≈ 𝑦2 − 𝑦 𝑦2 − 𝑦1 𝑓(𝑥, 𝑦1 ) + 𝑦…
Figure 7
Figure 7. Figure 7: Structure of license plate vertex estimation subnetwork CBL K(3,3) S(1,1) 64×22×92 Maxpool K(1,3,3) S(1,1,1) 64×20×90 Block 128×20×90 Maxpool K(1,3,3) S(2,1,2) 64×18×44 Block ×2 256×18×44 Maxpool K(1,3,3) S(4,1,2) 64×16×21 CBL K(1,4) S(1,1) 256×16×18 CBL K(13,1) S(1,1)…
Figure 8
Figure 8. Figure 8: Recognition model network structure diagram 3.3.1. LPRNet Optimization with LP-CA As shown in Figure8, LPRNet employs a multi-level feature fusion architecture, as shown in Figure9, high-level features focus more on global semantic information, while mid- and low-level…
Figure 9
Figure 9. Figure 9: Layer-wise Feature Visualization Recognition results 皖AA044 皖 A 0 4 - (a) [12,48,72,72,48,48,72,72,72,72,38,38,72,72,42,72,72,42] ①② ③ ① ③ ④ ④ Recognition results 皖AA044 皖 A 0 4 - (a) [12,48,72,72,48,48,72,72,72,72,38,38,72,72,42,72,72,42] ①② ③ ① ③ ④ ④ (b) [PITH_FULL_…
Figure 10
Figure 10. Figure 10: Visualization results of dictionary page feature map and CTC output character ID (a) Visualization results of dictionary page feature map (b) Visualization results of CTC output character ID for each character position, LPRNet performs column-wise average pooling on t…
Figure 11
Figure 11. Figure 11: LP-CA network structure [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: CCPD License Plate Distribution by Province In the context of license plate recognition tasks, traditional negative samples are absent. To address this, we adapt the standard Focal Loss approach for positive samples in the calculation of our Focal CTC Loss. This metho…
Figure 13
Figure 13. Figure 13: Main flowchart for building a double-line dataset calculation formula is as follows: Focal CTC loss = 𝛼(1 − 𝑝) 𝛾 ∗ CTC Loss (7) Focal CTC Loss enables license plate recognition networks to focus on challenging samples, thereby mitigating the adverse effects of class i…
Figure 14
Figure 14. Figure 14: Example of a generated double-line license plate Sampling blur Gaussian blur Horizontal motion blur Pixel blur Vertical motion blur Original figure Sampling blur Gaussian blur Horizontal motion blur Pixel blur Vertical motion blur Original figure [PITH_FULL_IMAGE:fig…
Figure 15
Figure 15. Figure 15: Various blurring effects 4. Experiment settings and Results Analysis The dataset used in this experiment is the publicly available CCPD license plate dataset. This dataset encompasses license plate images captured in complex environmental conditions, including various…
Figure 16
Figure 16. Figure 16: Examples of Labeling Errors in CCPD [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]
Figure 17
Figure 17. Figure 17: Examples of Selected Mislabeled License Plate Images [PITH_FULL_IMAGE:figures/full_fig_p015_17.png]
Figure 18
Figure 18. Figure 18: Sample Images from the Two-Line License Plate Dataset top-left and bottom-right corner coordinates of the license plate bounding box in the original labels to randomly perturb these positions. Additionally, the coordinates of each of the four vertices of the license p…
Figure 19
Figure 19. Figure 19: Samples of CCPD dataset Base other -test Base-train(79999) Base-val(19999) Base-test(100000) Challenge(10006) Db(20001) Fn(19999) Rotate(9998) Tilt(10000) Weather(9999) Base other -test Base-train(79999) Base-val(19999) Base-test(100000) Challenge(10006) Db(20001) Fn(…
Figure 20
Figure 20. Figure 20: Distribution Map of CCPD Subsets and License Plate Types [PITH_FULL_IMAGE:figures/full_fig_p016_20.png]
Figure 21
Figure 21. Figure 21: Samples of LSV dataset the proposed perturbation method provides favorable conditions for evaluating license plate recognition networks independently of localization networks. Page 16 of 28 [PITH_FULL_IMAGE:figures/full_fig_p016_21.png]
Figure 22
Figure 22. Figure 22 [PITH_FULL_IMAGE:figures/full_fig_p017_22.png]
Figure 23
Figure 23. Figure 23: License plate images obtained using different extraction methods: (a) Images based on ground truth labels, (b) Images based on YOLO detection results, (c) Images based on random coordinate perturbation of ground truth labels 4.2. LPTR-AFLNet Training The software plat…
Figure 24
Figure 24. Figure 24: Illustration of LPRTR-AFLNet Training Strategy [PITH_FULL_IMAGE:figures/full_fig_p018_24.png]
Figure 25
Figure 25. Figure 25: Samples of affine transformation rectification with STN (a)affine transformation rectification with STN (b)perspective transformation rectification with STN (c)perspective transformation rectification with PTR (a)affine transformation rectification with STN (b)perspec…
Figure 26
Figure 26. Figure 26: Comparison of rectification results of STN and PTR 4.3. Experimental results analysis 4.3.1. PTR performance analysis Traditional Spatial Transformer Networks (STN) typically regress only the affine transformation matrix, thereby limiting their capability to perform s…
Figure 27
Figure 27. Figure 27: Single-line license plate image rectification samples with PTR (a)double-line license plate images (b)rectification results with PTR [PITH_FULL_IMAGE:figures/full_fig_p020_27.png]
Figure 28
Figure 28. Figure 28: Double-line license plate image rectification samples with PTR 4.3.2. AFLNet Performance Analysis In this study, we conducted comparative experiments on the optimized model and mainstream license plate recognition methods using the CCPD dataset. The comparison results…
Figure 29
Figure 29. Figure 29: Sample images of license plates with recognition errors CCPD dataset. Specifically, Accuracy_7c and Accuracy_6c are improved by 25.90% and 22.84%, respectively, while CP increases by 6.69%. A statistical analysis was performed to evaluate the license plate image recog…
Figure 30
Figure 30. Figure 30: Bar Graph Visualization of error statistics for license plate types on the CCPD dataset A.皖AA044 B.皖AAA044 C.皖AAA044 [12,48,72,72,48,48,72,72,72,72,38,38,72,72,42,72,72,42] [12,48,48,72,72,48,48,72,48,48,72,38,38,72,72,42,72,42] [12,48,48,72,72,48,72,72,48,72,72,38,72…
Figure 31
Figure 31. Figure 31: Qualitative Error Analysis and Visualizations of Character Omission Errors dictionary page are attenuated. Consequently, even if the activation intensity of this particular ’A’ character is weaker than other ’A’ characters, it can still be correctly recognized. Figure…
Figure 32
Figure 32. Figure 32: Qualitative Error Analysis and Visualizations of Character Insertion Errors response for ‘1’ in the corresponding dictionary page within LPRNet, while the response for ‘-’ at that position is enhanced. As a consequence, the ‘-’ is correctly identified as a separator, …
Figure 33
Figure 33. Figure 33: Examples of incorrect license plate recognition in specific scenarios Finally, the fourth scenario, depicted in Figure33(d), stems from the poor quality of license plate images where the characters are blurred and even challenging to discern by the human eye. The indi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 24 canonical work pages

  1. [1]

    Improvinglicenseplaterecognitionviadiversestylisticplategeneration

    Liu,Q.,Chen,S.L.,Chen,Y.X.,Yin,X.C.,2024. Improvinglicenseplaterecognitionviadiversestylisticplategeneration. PatternRecognition Letters 183, 117–124

    work page 2024

  2. [2]

    Afa-net: Adaptive feature attention network in image deblurring and super-resolution for improving license plate recognition

    Kim, D., Kim, J., Park, E., 2024. Afa-net: Adaptive feature attention network in image deblurring and super-resolution for improving license plate recognition. Computer Vision and Image Understanding 238, 103879

    work page 2024

  3. [3]

    Real-timelicenseplatedetectionandrecognitionsystemusingyolov7xandeasyocr,in:2023GlobalConference on Information Technologies and Communications (GCITC), IEEE

    Dhyani,S.,Kumar,V.,2023. Real-timelicenseplatedetectionandrecognitionsystemusingyolov7xandeasyocr,in:2023GlobalConference on Information Technologies and Communications (GCITC), IEEE. pp. 1–5

    work page 2023

  4. [4]

    Deeplearningapproachtoautomaticallyrecogniselicensenumberplates

    Gautam,A.,Rana,D.,Aggarwal,S.,Bhosle,S.,Sharma,H.,2023. Deeplearningapproachtoautomaticallyrecogniselicensenumberplates. Multimedia Tools and Applications 82, 31487–31504

    work page 2023

  5. [5]

    Two-step algorithm for license plate identification using deep neural networks

    Kundrotas, M., Janut˙enait˙e-Bogdanien˙e, J., Šešok, D., 2023. Two-step algorithm for license plate identification using deep neural networks. Applied Sciences 13, 4902

    work page 2023

  6. [6]

    Spatial transformer networks

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

    2015

  7. [7]

    Research on the algorithm of license plate recognition based on mpgan haze weather

    Zhang, W., Lu, J., Zhang, J., Li, X., Zhao, Q., 2022. Research on the algorithm of license plate recognition based on mpgan haze weather. IEICE TRANSACTIONS on Information and Systems 105, 1085–1093

    work page 2022

  8. [8]

    Multi-model fusion vehicle supervision system based on deep learning, in: 2023 5th International Conference on Artificial Intelligence and Computer Applications (ICAICA), IEEE

    Shao, P., Yang, Y., Xu, R., Gao, S., Zhang, S., Wang, J., 2023. Multi-model fusion vehicle supervision system based on deep learning, in: 2023 5th International Conference on Artificial Intelligence and Computer Applications (ICAICA), IEEE. pp. 583–590

    work page 2023

Show all 34 references
  1. [9]

    Spatial transform depthwise over-parameterized convolution recurrent neural network for license plate recognition in complex environment

    Deng, J., Wei, H., Lai, Z., Gu, G., Chen, Z., Chen, L., Ding, L., 2023. Spatial transform depthwise over-parameterized convolution recurrent neural network for license plate recognition in complex environment. Journal of Computing and Information Science in Engineering 23, 011010

    2023

  2. [10]

    Internimage: Exploring large-scale vision foundation models with deformable convolutions, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al., 2023. Internimage: Exploring large-scale vision foundation models with deformable convolutions, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

    2023

  3. [11]

    Irregularlicenseplaterecognitionviaglobalinformationintegration,in:International Conference on Multimedia Modeling, Springer

    Liu,Y.Y.,Liu,Q.,Chen,S.L.,Chen,F.,Yin,X.C.,2024. Irregularlicenseplaterecognitionviaglobalinformationintegration,in:International Conference on Multimedia Modeling, Springer. pp. 325–339

    2024

  4. [12]

    Real-time vehicle classification and license plate recognition via deformable convolution based yolo v8 network

    Srinivasan, R., Rajeswari, D., Arivarasi, A., Govindasamy, A., 2024. Real-time vehicle classification and license plate recognition via deformable convolution based yolo v8 network. IEEE Sensors Journal

    2024

  5. [13]

    A robust layout-independent license plate detection and recognition model based on attention method

    Seo, T.M., Kang, D.J., 2022. A robust layout-independent license plate detection and recognition model based on attention method. IEEE Access 10, 57427–57436

    2022

  6. [14]

    Unified chinese license plate detection and recognition with high efficiency

    Gong, Y., Deng, L., Tao, S., Lu, X., Wu, P., Xie, Z., Ma, Z., Xie, M., 2022. Unified chinese license plate detection and recognition with high efficiency. Journal of Visual Communication and Image Representation 86, 103541

    2022

  7. [15]

    Peng, W., Hu, Z., Liang, T., 2023. End-to-end efficient cascade license plate recognition system in unconstrained scenarios, in: 2023 IEEE 3rd International Conference on Power, Electronics and Computer Applications (ICPECA), IEEE. pp. 963–968

    2023

  8. [16]

    Improving multi-type license plate recognition via learning globally and contrastively

    Liu, Q., Liu, Y., Chen, S.L., Zhang, T.H., Chen, F., Yin, X.C., 2024. Improving multi-type license plate recognition via learning globally and contrastively. IEEE Transactions on Intelligent Transportation Systems 25, 11092–11102. doi:10.1109/TITS.2024.3365537

    2024

  9. [17]

    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 2022. arXiv preprint arXiv:2206.03001

    2022 arXiv

  10. [18]

    Lprnet: License plate recognition via deep neural networks

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

    2018 arXiv

  11. [19]

    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

    2022

  12. [20]

    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

  13. [21]

    SiLab-Bonn, . Ccpdv2. [Online].https://github.com/ultralytics/ultralytics

  14. [22]

    Sangctc-improving traditional ctc loss function for optical music recognition (august 2019), in: 2019 IEEE Pune Section International Conference (PuneCon), IEEE

    Shah, J.K., Padte, A.S., Ahirao, P.N., 2019. Sangctc-improving traditional ctc loss function for optical music recognition (august 2019), in: 2019 IEEE Pune Section International Conference (PuneCon), IEEE. pp. 1–5

    2019

  15. [23]

    Spatial correction and deblurring fusion algorithm for vehicle license plate images based on deep learning

    Ding, F., Zhang, D., 2024. Spatial correction and deblurring fusion algorithm for vehicle license plate images based on deep learning. International Journal of Modeling, Simulation, and Scientific Computing 15, 2450045

    2024

  16. [24]

    A robust license plate detection and recognition system based on detr and cnn

    Yang, C., Yu, S., Cai, C., 2022. A robust license plate detection and recognition system based on detr and cnn. Available at SSRN 4184430

    2022

  17. [25]

    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

  18. [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

  19. [27]

    12216–12224

    Wang,T.,Zhu,Y.,Jin,L.,Luo,C.,Chen,X.,Wu,Y.,Wang,Q.,Cai,M.,2020.Decoupledattentionnetworkfortextrecognition,in:Proceedings of the AAAI conference on artificial intelligence, pp. 12216–12224

    2020

  20. [28]

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

    2020

  21. [29]

    Flexiblecameracalibrationbyviewingaplanefromunknownorientations,in:Proceedingsoftheseventhieeeinternational conference on computer vision, Ieee

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

    1999

  22. [30]

    Graves, A., Fernández, S., Gomez, F., Schmidhuber, J., 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks, in: Proceedings of the 23rd international conference on Machine learning, pp. 369–376

    2006

  23. [31]

    Focal loss for dense object detection, in: Proceedings of the IEEE international conference on computer vision, pp

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P., 2017. Focal loss for dense object detection, in: Proceedings of the IEEE international conference on computer vision, pp. 2980–2988

    2017

  24. [32]

    Accurate license plate location based on synchronous vertex and body region detection

    XU, G., LIU, G., KUANG, W., WAN, Q., MA, G., LEI, B., 2022. Accurate license plate location based on synchronous vertex and body region detection. Journal of Beijing University of Aeronautics and Astronautics 50, 376–387. Page 27 of 28

    2022

  25. [33]

    An efficient and layout-independent automatic license plate recognition system based on the yolo detector

    Laroca, R., Zanlorensi, L., Gonçalves, G., Todt, E., Schwartz, W., Menotti, D., 2021. An efficient and layout-independent automatic license plate recognition system based on the yolo detector. arxiv 2019. arXiv preprint arXiv:1909.01754

    2021 arXiv

  26. [34]

    Anend-to-endtrainableneuralnetworkforimage-basedsequencerecognitionanditsapplicationtoscenetext recognition

    Shi,B.,Bai,X.,Yao,C.,2016. Anend-to-endtrainableneuralnetworkforimage-basedsequencerecognitionanditsapplicationtoscenetext recognition. IEEE transactions on pattern analysis and machine intelligence 39, 2298–2304. Page 28 of 28

    2016

Pith tools

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