license plate recognition

To address the limited robustness of Automatic License Plate Recognition (ALPR) systems caused by significant cross-national variations in license plate layouts and strong real-world noise, this paper proposes a pattern-aware end-to-end license plate recognition framework. Departing from explicit layout classification and hand-crafted rules, our approach implicitly encodes structural priors of license plates through joint optimization, integrating a high-precision detection network, a vision Transformer, and an iterative language modeling module. This enables seamless integration of character recognition and semantic post-processing. Evaluated on international benchmarks—including IR-LPR, UFPR-ALPR, and AOLP—our method substantially outperforms existing segmentation-free approaches. It achieves state-of-the-art accuracy and generalization under challenging conditions such as severe geometric distortion, low resolution, non-standard fonts, and cluttered backgrounds.

To address the high computational overhead and poor real-time performance of Automatic License Plate Recognition (ALPR) systems in video streams—largely caused by multi-frame dependency—this paper proposes a single-frame-driven efficient ALPR method. Our approach introduces (1) Visual Rhythm (VR) modeling to encode vehicle motion trajectories into spatiotemporal images, and (2) an Accumulated Line Analysis (ALA) algorithm that jointly performs license plate localization and character sequence extraction within a single frame. The framework integrates YOLOv5-based detection, a lightweight CNN for character recognition, and VR-ALA joint inference. Evaluated on real-world traffic video datasets, our method achieves competitive accuracy (>92%) compared to state-of-the-art multi-frame approaches, while improving processing speed by 3.1× and significantly reducing GPU memory consumption and end-to-end latency. To the best of our knowledge, this is the first end-to-end ALPR system operating at the single-frame and per-vehicle granularity, establishing a new paradigm for edge deployment.

This study addresses the challenges of automatic license plate recognition for Bengali due to its complex character set and irregular layout, which hinder accurate localization and text extraction. The authors propose an end-to-end recognition system that first enhances localization precision through a two-stage adaptive training strategy based on YOLOv8, achieving a localization accuracy of 97.83% and an IoU of 91.3%. Subsequently, they introduce a novel vision-language OCR model that uniquely integrates Vision Transformer with BanglaBERT to enable sequential text generation. This approach demonstrates superior performance with a character error rate of 0.1323 and a word error rate of 0.1068, significantly improving robustness across diverse real-world scenarios and proving effective in complex, unconstrained environments.

To address the challenges of recognizing diverse Chinese and English license plates under open-world conditions—characterized by complex imaging environments and scarce annotated bilingual plate data—this paper proposes a lightweight vision-language collaborative framework. Methodologically, it integrates a lightweight Vision Transformer (ViT) visual encoder with a sequence transcription decoder, and introduces a novel perspective correction module jointly supervised by corner-point regression and viewpoint classification to enhance robustness and interpretability. Additionally, synthetic data augmentation, texture-mapping-based realism enhancement, and a coordinate-regression auxiliary network are incorporated to reduce annotation dependency. Evaluated on the CCPD dataset, the framework achieves 99.34% recognition accuracy under coarse localization interference, 99.58% under precise localization, and 98.70% for bilingual plates, operating at 167 FPS—demonstrating both high accuracy and practical efficiency.

To address the joint bottlenecks of accuracy, real-time performance, and edge deployability in license plate detection and recognition (LPR) systems under complex traffic scenarios, this paper proposes a lightweight two-stage YOLOv8 collaborative framework: YOLOv8n for efficient license plate localization and YOLOv8s for character recognition. We introduce a novel RNN-free and attention-free character serialization algorithm based on X-axis coordinates. Integrated with multi-scale data augmentation and edge-optimized inference, the system achieves Precision = 0.964 and mAP₅₀ = 0.918 for plate detection, and Precision = 0.92 and mAP₅₀ = 0.91 for character recognition—outperforming existing lightweight LPR methods. Crucially, it supports real-time inference on edge devices. Our core contributions are: (1) a dual-model collaborative architecture tailored for LPR, and (2) a lightweight, geometry-driven character ordering paradigm that eliminates sequential modeling overhead.

This study addresses the challenges of low license plate recognition accuracy and fragmented tracking trajectories in dynamic traffic surveillance, caused by abrupt illumination changes, extreme viewing angles, high-speed motion, and occlusions. To tackle these issues, the authors propose a five-stage end-to-end pipeline that leverages YOLOv8-nano for joint vehicle and license plate detection, integrates the SORT algorithm for multi-object tracking, and introduces an innovative offline temporal bounding box interpolation mechanism to recover broken trajectories. Furthermore, the framework incorporates location-guided OCR (EasyOCR) fused with a license plate syntax validation module to enhance recognition accuracy and spatiotemporal consistency under complex conditions. Experimental results demonstrate that the proposed approach significantly improves both the continuity of license plate tracking and the robustness of recognition in highly challenging dynamic scenarios.

This work addresses the challenge of low-cost, real-time license plate recognition in complex, unstructured traffic environments typical of developing countries by proposing a lightweight two-stage system tailored for embedded platforms. The approach leverages a compact convolutional neural network for both license plate detection and character recognition. Key contributions include the creation of SL-LPR, the first dataset specifically designed for such scenarios, and an efficient deployment on the Xilinx Kria KV260 FPGA using Brevitas for low-bit quantization and the FINN framework. Evaluated on the SL-LPR dataset, the system achieves a detection mAP of 93.6% and a character recognition accuracy of 87.88%, delivering end-to-end inference at 11.5 frames per second.

This study addresses the challenges of license plate detection and recognition (LPDR) in data-scarce regions with visually distinctive plates, such as Bolivia, where performance is often degraded by perspective distortion and illumination variations. The authors propose a two-stage robust recognition framework: first, a YOLO-based detector is pre-trained on synthetic data generated via Blender to simulate extreme imaging conditions, followed by domain-adaptive fine-tuning using street-view imagery from La Paz. Detected plates undergo geometric rectification before being fed into a character recognition model, with a lightweight vision-language model (Gemma3-4B) serving as a fallback mechanism triggered by low prediction confidence. This work introduces the first public LPDR dataset for Bolivian license plates and achieves a character-level accuracy of 89.6% on real-world street scenes, significantly enhancing recognition robustness in complex urban environments.

This work addresses the challenge of scale ambiguity in monocular visual odometry, which often leads to insufficient metric accuracy. The authors propose an unsupervised method that requires neither training data nor expensive sensors, leveraging the standardized geometric layout of U.S. license plates as a novel geometric prior to enable metric-scale distance estimation. The system integrates parallel quadruple license plate detection, a three-stage state verification pipeline (combining OCR matching, color scoring, and a lightweight neural network), and inverse-variance-weighted deep fusion, followed by a one-dimensional constant-velocity Kalman filter to produce smooth estimates of distance, relative velocity, and collision warnings. Experiments demonstrate an average absolute error of 2.3% at 10 meters, a 36% reduction in distance variance compared to the conventional plate-width method, and relative error five times lower than deep learning baselines, while maintaining robust and continuous output even under brief occlusions.

This work addresses the challenge of deploying efficient automatic license plate recognition (ALPR) on resource-constrained microcontroller units (MCUs) without relying on dedicated hardware accelerators. Leveraging a 9-core RISC-V GAP8 processor and an ultra-low-power grayscale image sensor, the authors implement a full-pipeline edge-based ALPR system that integrates SSDlite-MobilenetV2 for license plate detection and LPRNet for character recognition. Through synergistic model compression and multi-core scheduling optimizations, this study achieves the first end-to-end deployment of multiple deep learning models on an MCU-class platform. Experimental results demonstrate a detection mAP of 38.9% and character recognition accuracy exceeding 99.13% on public datasets, with the capability to recognize plates as small as 30×5 pixels. The system operates at a power consumption of only 117 mW, achieving a 73× improvement in energy efficiency compared to the Raspberry Pi 3.