GERIS: A Game-Theoretic Framework for Filtering Instance-Dependent Label Noise in License Plate Data Augmentation

Seyedeh Sara Jalili Shani; Rouhollah Ahmadian; Amin Rahmani; Mahdi Bideh; Mehdi Ghatee

cs.CVSep 22, 2026

GERIS: A Game-Theoretic Framework for Filtering Instance-Dependent Label Noise in License Plate Data Augmentation

Authors: Seyedeh Sara Jalili Shani, Rouhollah Ahmadian, Amin Rahmani, Mahdi Bideh, Mehdi Ghatee

Organizations: Department of Computer Science, University of Alberta, Alberta, Canada · Department of Computer Science, University of Alberta Edmonton, Alberta, Canada · Department of Mathematics and Computer Science, Amirkabir University of Technology, Iran · Department of Mathematics & Computer Science, Amirkabir University of Technology Tehran, Iran

Abstract

In this paper, we propose GERIS, a game-theoretic framework for instance selection in the data augmentation phase of license plate recognition systems. During augmentation, synthetic license plate images are generated and transformed using stochastic noise to simulate real-world conditions. However, certain noise configurations lead to highly distorted, unreadable images that degrade model performance by introducing instance-dependent label noise. GERIS formulates a non-cooperative game in which each noise vector competes for inclusion in the training set based on its similarity to labeled data and its contribution to model reliability. By identifying and pruning low-quality instances, GERIS improves the overall quality of the augmented dataset. Unlike traditional black-box learning methods, GERIS offers a transparent, theoretically grounded mechanism for data filtering. Experimental results demonstrate that GERIS outperforms existing instance selection methods in terms of classification accuracy and robustness.

Figures & tables

Instance Selection Methods			Description	Ref.
Filtering-based Instance Selection	Distance-based filtering selects instances by their proximity to others.	DBSCAN	A density-based clustering algorithm groups data points without requiring a preset cluster count, selecting representatives from dense regions.	[ 2 ]
		K-nearest Neighbors (KNN)	KNN classifies data points by their nearest neighbors’ majority class. K-means clusters data by similarity, selecting centroids as representatives.	[ 13 ]
		Isolation Forest	An isolation forest detects anomalies by isolating rare, distinct instances using random forests.	[ 14 ]
		K-means	K-means clusters data into k groups by minimizing squared distances to centroids, using them as representatives.	[ 15 ]
		Condensed Nearest Neighbor	A method that iteratively selects only correctly classified, essential instances for an accurate reduced dataset.	[ 16 ]
	Class distribution filtering selects instances to represent the dataset’s class diversity.	Edited Nearest Neighbor (ENN)	A method that removes misclassified or redundant instances by comparing them to their nearest neighbors, boosting model performance.	[ 17 ]

Table 1: A review of algorithms used for instance selection.

Figure 1: The architecture of generating license plate.

Transformation	Description
Salt and Pepper	The Salt and Pepper noise is a simple noise that simulates dead pixels in a camera and digital noise. This noise uses two values corresponding to salt and pepper, respectively, $v_{s}$ and $v_{p}$ .
Blur	The Blur noise uses Gaussian and median filters to add a blur effect to the images. This noise uses two parameters, filter size, and variance. An increase in filter size or variance will increase the blur.
Circular Light	The circular light noise is the simulation of the reflection of light in some images in the form of fuzzy circles. To obtain such noise, first, we apply a median filter to smooth it out and prevent unwanted circular lights. After that, we convert the image format from RGB to YUV and add noise to its Y channel. The noise is a Gaussian kernel amplified by a factor $l_{p}$ .
Light	The light noise is the replica of high exposure in images that adds extra brightness to an image. This phenomenon makes text to be less readable. To construct such noise, we proceed as the Circular light noise and convert the image format to YUV after applying median smoothing. Following these steps, we add light noise as a constant to the Y channel of the image.
Gradient Light	The gradient light noise is another light noise that can produce noises similar to light beams. This noise occurs when a light source is in the environment and the camera’s lens is not clean. The initial steps for producing this noise are similar to other light noises. After obtaining YUV channels, we select a linear path and add a constant to the Y channel. In the last step, we convert the image to RGB channels and obtain a noisy image.
Negative	The negative noise works the opposite way of the light noises. This noise lowers the exposure and light in the image by deducting a constant from the Y channel of the YUV format of the image. YUV format of the image and reverse conversions are similar to light noises, and the only difference is the subtraction part.

Table 2: Transformation Functions

Figure 2: Four samples of correct and corrupt noise vectors. (a) refers to the correct samples, and (b) refers to the corrupt samples.

Figure 3: Two stages of evaluation methodology.

Panel of Figure 4 — Figure 4: Analysis of training size ratio in terms of mean F1 score.

Method	F1	$\mathrm{\textbf{Precision}}_{\mathrm{\textbf{corrupt}}}$	$\textbf{Precision}_{\textbf{correct}}$	$\textbf{Recall}_{\textbf{corrupt}}$	$\textbf{Recall}_{\textbf{correct}}$
KNN	73.61 ( $\pm$ 0.002)	48.62 ( $\pm$ 0.000)	97.72 ( $\pm$ 0.000)	48.69 ( $\pm$ 0.000)	99.48 ( $\pm$ 0.0)
MLP	73.28 ( $\pm$ 0.003)	49.54 ( $\pm$ 0.000)	97.04 ( $\pm$ 0.000)	46.85 ( $\pm$ 0.000)	99.84 ( $\pm$ 0.000)
Naïve Bayes	78.49 ( $\pm$ 0.015)	55.07 ( $\pm$ 0.000)	99.15 ( $\pm$ 0.000)	84.27 ( $\pm$ 0.002)	92.11 ( $\pm$ 0.0)
SVM	73.67 ( $\pm$ 0.003)	49.16 ( $\pm$ 0.000)	97.61 ( $\pm$ 0.000)	48.27 ( $\pm$ 0.000)	99.69 ( $\pm$ 0.000)
Rand Forest	73.80 ( $\pm$ 0.002)	49.05 ( $\pm$ 0.000)	97.76 ( $\pm$ 0.000)	48.78 ( $\pm$ 0.000)	99.64 ( $\pm$ 0.000)
One Class SVM	56.45 ( $\pm$ 0.023)	29.60 ( $\pm$ 0.002)	87.06 ( $\pm$ 0.001)	39.63 ( $\pm$ 0.003)	87.75 ( $\pm$ 0.001)

Table 3: Benchmark of Filtering Methods (Mean

\pm

Std over 100 Trials)

Method	Run Time (s)	Memory Usage (MB)
KNN	2.09 ( $\pm$ 0.58)	212.38 ( $\pm$ 18.29)
MLP	6.18 ( $\pm$ 1.58)	194.34 ( $\pm$ 26.57)
Naïve Bayes	3.07 ( $\pm$ 0.05)	183.33 ( $\pm$ 5.65)
SVM	2.13 ( $\pm$ 0.05)	210.71 ( $\pm$ 8.67)
Rand Forest	1.80 ( $\pm$ 0.02)	241.83 ( $\pm$ 9.57)
One Class SVM	2.17 ( $\pm$ 0.06)	203.37 ( $\pm$ 21.27)

Table 4: Resource Usage Profile of Filtering Methods (Mean

\pm

Std over 100 Trials)

Training Data		Filtering Method	Digits				Alphabets
Real Data	Synthetic	Filtering Method	Accuracy	F1	Precision	Recall	Accuracy	F1	Precision	Recall
✓	✗	✗	87.23	79.79	79.56	80.99	70.5	57.75	62.25	58.32
✗	✓	✗	94.95	94.84	95.67	94.5	92.71	82.61	82.53	87.79
✗	✓	Naive Bayes	96.21	96.21	96.4	96.07	94.52	86.62	86.91	88.28
✓	✓	Naive Bayes	96.21	96.21	96.4	96.07	95.75	87.61	86.8	90.59
✗	✓	GERIS	96.64	96.26	96.67	95.93	96.41	87.98	89.06	88.04

Table 5: Benchmark of Character Recognition

Game-Theoretic RL

See all 28 papers

LPLCv2: An Expanded Dataset for Fine-Grained License Plate Legibility Classification
Apr 9, 2026Lucas Wojcik, Eduardo A. F. Machoski, Eduil Nascimento +2Automatic License Plate Recognition
Federal University of Paran´a, Curitiba, Brazil · Department of Technological Development and Quality, Paran´a Military Police, Curitiba, Brazil · Pontifical Catholic University of Paran´a, Curitiba, Brazil
Next-Generation Parallel Decoder for LPDR: Architectural Optimization and Class-Balanced GAN-Augmentation
Jun 4, 2026Shawaiz Obaid, Nida Chandio, Neha Jamil +1Class-Imbalanced Learning Synthetic Data Augmentation
School of Electrical Engineering and Computer Science National University of Sciences & Technology Islamabad, Pakistan
MoireMix: A Formula-Based Data Augmentation for Improving Image Classification Robustness
Mar 26, 2026Yuto Matsuo, Yoshihiro Fukuhara, Yuki M. Asano +3Data Augmentation Synthetic Data Augmentation
National Institute of Advanced Industrial Science and Technology (AIST) · Tokyo Denki University · Waseda University +2

Apr 9, 2026J/K move · Enter open · S save

LPLCv2: An Expanded Dataset for Fine-Grained License Plate Legibility Classification

Lucas Wojcik, Eduardo A. F. Machoski, Eduil Nascimento, Rayson Laroca+1

Federal University of Paran´a, Curitiba, Brazil · Department of Technological Development and Quality, Paran´a Military Police, Curitiba, Brazil · Pontifical Catholic University of Paran´a, Curitiba, Brazil

Automatic License Plate Recognition

Open paper AskPDF arXiv

Modern Automatic License Plate Recognition (ALPR) systems achieve outstanding performance in controlled, well-defined scenarios. However, large-scale real-world usage remains challenging due to low-quality imaging devices, compression artifacts, and suboptimal camera installation. Identifying illegible license plates (LPs) has recently become feasible through a dedicated benchmark; however, its impact has been limited by its small size and annotation errors. In this work, we expand the original benchmark to over three times the size with two extra capture days, revise its annotations and introduce novel labels. LP-level annotations include bounding boxes, text, and legibility level, while vehicle-level annotations comprise make, model, type, and color. Image-level annotations feature camera identity, capture conditions (e.g., rain and faulty cameras), acquisition time, and day ID. We present a novel training procedure featuring an Exponential Moving Average-based loss function and a refined learning rate scheduler, addressing common mistakes in testing. These improvements enable a baseline model to achieve an 89.5% F1-score on the test set, considerably surpassing the previous state of the art. We further introduce a novel protocol to explicitly addresses camera contamination between training and evaluation splits, where results show a small impact. Dataset and code are publicly available at https://github.com/lmlwojcik/LPLCv2-Dataset.

Abstract

Game-Theoretic RL

Explore similar work

LPLCv2: An Expanded Dataset for Fine-Grained License Plate Legibility Classification

Next-Generation Parallel Decoder for LPDR: Architectural Optimization and Class-Balanced GAN-Augmentation

MoireMix: A Formula-Based Data Augmentation for Improving Image Classification Robustness