Style-Preserving Generator
for Synthetic License Plate Recognition
(TCSVT)

The training of robust License Plate Recognition (LPR) models has long been constrained by the inherent limitations of real-world data. These datasets often suffer from significant data imbalance, where attributes like character frequency, illumination, and plate orientation are unevenly distributed, leading to biased models with degraded performance. Furthermore, the use of real license plates raises substantial privacy concerns, as each plate is directly linked to its owner.
This research introduces the Style-Preserving Generator (SPG), a novel generative AI framework designed to address these challenges. The SPG synthesizes highly realistic license plate images by editing the characters on existing plates while meticulously preserving their original visual styles, including lighting, texture, and orientation. This approach allows for the creation of balanced and anonymized datasets. The paper’s contributions are twofold: the development of the SPG model and the introduction of LP-2025, a new, challenging benchmark dataset created to validate the SPG's efficacy and spur further advancements in LPR research.

The framework of our proposed method primarily includes a transformer T, an encoder Ce, a decoder Cd, a target generator G, and a discriminator Dt . Ce converts the source image Is into a latent code cs. Subsequently, the transformer T takes the latent code cs and the word embeddings of the target string st as inputs to generate an attention feature code ft at, which is then decoded by Cd to estimate the target character mask image Bt , as illustrated by the green path. The blue path demonstrates the process of estimating the source character mask Bs for the source string ss . Finally, G is responsible for generating the desired target license plate It.

The LP-2025 dataset consists of 67,429 license plates (LPs) found in 17,789 images taken in the Taipei metropolitan area over 6 months under various conditions.
Of the total LPs, 33,240 are classified as "readable" for tasks like License Plate Recognition (LPR), and 34,189 are "unreadable" (too small or blurry). The complete dataset, including both readable and unreadable plates, is suitable for studies in License Plate Detection (LPD), generation, or deblurring.

Key characteristics of the dataset include:

• High Variation: LPs range in size from 4×8 to 1850×1774 pixels and have wide orientation variations (yaw: −75∘∼75∘, pitch: −30∘∼60∘, roll: −50∘∼50∘).

• High Density: Uniquely, 60% of the images contain more than two LPs, with an average of 1.87 LPs per image. Each LP is annotated with its four corner coordinates.

To address the natural imbalance of hard-to-recognize characters, the dataset was augmented with synthetically generated characters.

The data is split into three subsets, ensuring a similar distribution of attributes:

• Training (40%): 13,243 readable and 13,638 unreadable LPs.

• Validation (10%): 3,340 readable and 3,416 unreadable LPs.

• Testing (50%): 16,657 readable and 17,135 unreadable LPs.

The source code can be downloaded from here.

Click here for more information.