Style-Preserving Generator for Synthetic License Plate Recognition
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
We propose the Style-Preserving Generator (SPG) to generate synthetic license plate data to train License Plate Recognition (LPR) models, and compare the performance with the same models trained on real-world data. The proposed SPG can edit the characters on real-world license plates while maintaining their original styles, allowing synthetic license plate data to be generated with user-specified characters. We can therefore synthesize license plates with desired characters to effectively alleviate the data attribute imbalance and privacy issues associated with real-world license plates. To the best of our knowledge, this work is the first study to present the making of synthetic LP data by proposing a novel text-editing approach tailor-made for LP data, that is the proposed SPG. The SPG consists of a transformer, a source encoder, a source style encoder, a character mask decoder, a target generator, and a target discriminator. Given a source license plate image and a specified text as input, these components collaborate to compute the self- and cross-attention embeddings, predict character masks, and generate a synthetic license plate in the source style but with source characters replaced by the specified characters. We adopt a two-phase training scheme. Phase 1 training uses synthetic data only, but Phase 2 training uses synthetic and real-life data. To showcase the effectiveness of the SPG, we introduce a new benchmark dataset, the LP-2025 (License Plate 2025), which alleviates the limitations of existing datasets and presents new challenges for license plate recognition and generative models. We validate SPG performance on the LP-2025 dataset and other benchmark datasets and compare it against state-of-the-art text-editing approaches.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it