Source-linked AI summary
A Cloud-Based Hybrid Model for Real-Time Detection of BRTA-Approved Licence Plates Using YOLO Tiny and Haar Cascade
Debashis Kar Suvra, Tahsina Farah Sanam
TL;DR
Bengali BRTA license plate detection is difficult because localized scripts and changing real-world conditions challenge conventional systems. The paper combines YOLO Tiny v4, Haar-Cascade, dynamic retraining, and a public dataset, reporting 94% precision, 93% recall, and a 93% F1 score.
Problem
Bengali license plate recognition lacks robust publicly available datasets and systems that handle lighting variation, motion blur, occlusion, and changing conditions.
Method
The paper combines YOLO Tiny v4 vehicle detection, Haar-Cascade plate localization, dynamic retraining, and a custom dataset of BRTA-compliant plates.
Results
94% precision, 93% recall, and a 93% F1 score were achieved by the hybrid model.
Takeaways & Limitations
The framework provides a resource-conscious Bengali license plate detection approach for varied environmental conditions and supports further non-Latin script recognition research.
Abstract
from arXiv · showhide
Accurate vehicle license plate detection is essential for applications such as intelligent transportation systems, toll collection, parking management, and law enforcement. In Bangladesh, this task presents distinct challenges due to the complexity of localized license plates and environmental factors like lighting, occlusion, motion blur, and obstructions such as dirt or mud. These challenges often render conventional methods ineffective. This paper introduces a novel hybrid approach, combining the YOLO Tiny deep learning model with the Haar-Cascade classifier, for enhanced detection and localization of Bengali license plates. A key innovation of our system is the integration of a dynamic retraining pipeline, which allows the model to adapt to evolving real-world conditions. This retraining mechanism significantly boosts performance in low-confidence scenarios by continuously improving the model's accuracy as new data is encountered. Additionally, a publicly accessible dataset of BRTA-compliant license plates, captured under diverse and challenging conditions, has been developed to support this approach. Experimental results demonstrate that our approach not only achieves superior detection accuracy and computational efficiency over conventional models but also ensures consistent performance in resource-constrained environments, particularly in Bangladesh.
I. INTRODUCTION
Bengali BRTA plates challenge conventional recognition because localized scripts, varied designs, and real-world conditions reduce detection reliability. The paper proposes a hybrid YOLO Tiny v4–Haar-Cascade system with dynamic retraining and contributes a public dataset for these conditions.
- Bengali script, distinct plate designs, lighting variation, occlusion, motion blur, and vehicle diversity make conventional license plate recognition inaccurate.
- The proposed method combines YOLO Tiny v4 for efficient vehicle detection with Haar-Cascade for precise plate localization inside vehicle bounding boxes.
- Dynamic retraining uses low-confidence detections to adapt the model to new conditions and maintain accuracy in rapidly changing environments.
- Existing research lacks publicly accessible Bengali license plate datasets and adaptive systems that continuously learn from changing conditions.
- The study introduces a public Bengali license plate dataset captured under varied real-world conditions to support future localized-script recognition research.
II. LITERATURE REVIEW
The literature traces a transition from traditional Haar-cascade and HOG methods to CNN-based detection, while positioning computer vision and cloud computing as foundations for scalable visual analysis.
- Traditional Haar cascades and HOG established early object detection, but limitations in complex scenes and multiple classes encouraged CNN-based approaches.
- Computer vision uses algorithms to extract and analyze visual features from images and videos, supporting applications such as surveillance, medical imaging, and robotics.
- Cloud computing provides on-demand scalable resources that can support computationally demanding computer vision applications without maintaining local infrastructure.
III. PROPOSED METHODOLOGY
The methodology combines transfer learning, YOLOv4-tiny vehicle detection, Haar-Cascade plate localization, and real-time video analysis. Its custom dataset uses varied capture conditions and vehicle types to support robustness.
- The system uses transfer learning with YOLOv4-tiny for vehicle detection and Haar-Cascade for Bengali license plate localization.
- A. Data Collection & Dataset Development: The custom dataset contains vehicle images and videos collected with an IP camera and mobile phone to address Bengali license plate data scarcity.
- A. Data Collection & Dataset Development: Dataset videos cover bright sunlight, moderate daylight, and low-light night scenes, while images also include clear and rainy weather.
- A. Data Collection & Dataset Development: The dataset includes cars, motorcycles, auto-rickshaws, vans, and buses to broaden coverage across vehicle classes and road conditions.
B. YOLOv4 tiny for Vehicle Detection
YOLOv4-tiny performs real-time vehicle detection by predicting bounding boxes, confidence scores, and class probabilities for each grid cell. Postprocessing removes detections below a confidence threshold.
- YOLOv4-tiny predicts vehicle bounding boxes and confidence scores for each grid cell during real-time object detection.
- Each predicted box contains center coordinates, width, height, and a confidence score indicating whether it contains an object.
- The model estimates the probability that a bounding box contains a particular class using class-specific weights and the concatenated box representation.
- A confidence threshold discards low-confidence bounding boxes during postprocessing to filter unreliable detections.
C. Haar-Cascade for License Plate Localization
The Haar-Cascade classifier localizes BRTA-aligned license plates using learned Haar-like features, weights, and thresholds to distinguish positive plate regions from negative samples.
- The classifier is trained on positive BRTA-aligned license plate regions and negative regions without license plates.
- Each Haar-like feature receives a learned weight and threshold during classifier training.
- The decision function H(x, y) combines feature responses with learned weights and compares them against a threshold.
- Haar-like features compute differences between pixel-intensity sums in white and black rectangular regions.
- After detection, the system stores the license-plate result only when its highest confidence rate exceeds 80%.
D. Fine Tuning, Transfer learning & Hybrid Approach
The system fine-tunes YOLOv4-Tiny for Bengali license plates while retaining pretrained feature-extraction layers, then combines it with Haar-Cascade localization.
- Transfer learning retains YOLOv4-Tiny’s earlier feature-extraction layers and fine-tunes its detection head on Bengali license-plate data.
- Keeping early layers frozen preserves learned features and reduces computational demands during adaptation.
- The training setup selects anchors for license-plate aspect ratios and uses batch size 16 with a learning rate of 0.001.
1) Efficiency Strategies:
The efficiency strategy concentrates license-plate processing on vehicle regions of interest after initial vehicle detection, reducing computation on distant or irrelevant objects.
- The pipeline focuses computational resources on vehicles closer to the camera, around 1 meter or more.
- After YOLOv4 identifies vehicles across the frame, license-plate processing is restricted to extracted regions of interest.
- Using YOLOv4 for initial detection and restricting later processing to the ROI conserves resources and improves responsiveness.
2) Cascade Approach and Elimination of Irrelevant Regions:
The cascade processes video frames sequentially: YOLOv4 detects vehicles, then Haar-Cascade localizes plates within detected vehicle regions in a cloud-hosted real-time setup.
- YOLOv4 performs broad vehicle detection before Haar-Cascade examines license plates in the detected regions.
- This sequential processing avoids unnecessary computation on irrelevant regions and contributes to Bangla license-plate detection accuracy.
- A Flask server hosted the hybrid model in the cloud, while RTSP streams simulated live traffic for real-time evaluation.
- The Haar-Cascade classifier acts as a wake-up mechanism after vehicle detection and outputs the vehicle with its corresponding BRTA-aligned plate.
- The testing setup provided a scalable environment for evaluating the hybrid model under real-world conditions.
IV. EXPERIMENTAL STUDY
The experimental study evaluates the hybrid license-plate detection system using real-world BRTA-compliant Bengali plate data, standardized video processing, and both local and cloud-based real-time testing. Evaluation covers recognition metrics and streaming performance under constrained deployment conditions.
- The evaluation measures precision, recall, F1 score, accuracy, and frames per second on BRTA-compliant Bengali plates captured under real-world conditions.
- Experiments were conducted on an AMD Ryzen 3 2200G system with 16GB RAM to prioritize accessible real-time inference.
- Test videos were standardized to 464x464 pixels before processing in local and cloud-based environments.
- Live RTSP footage from a Dahua IMOU CRUISER SE+4MP camera was used for hybrid-model testing with confidence and suppression thresholds.
- License plates were localized with Haar Cascade using a 1.05 scaling factor, five minimum neighbors, and a minimum size of 60x80 pixels.
1) Data Augmentation Techniques:
The study examines how augmentation and selective transfer learning affect the hybrid model under lighting variation and occlusion. Results are assessed across recognition quality and processing speed, with Figure 3 comparing augmented and non-augmented models.
- Data Augmentation Techniques: Data augmentation targets lighting and signal variation through scaling, contrast adjustment, dynamic range enhancement, and noise reduction.
- Data Augmentation Techniques: The YOLOv4-Tiny detector was pre-trained on MS COCO, with only final prediction and classification layers fine-tuned because the dataset was limited.
- Data Augmentation Techniques: Selective fine-tuning improved adaptation to lighting variations and occlusions despite insufficient data for full retraining.
- Data Augmentation Techniques: The hybrid model was tested with live RTSP streams while Tables I and II report precision, recall, F1 score, accuracy, and FPS.
- Data Augmentation Techniques: Figure 3 compares performance for models trained with and without data augmentation techniques.
V. CONCLUSION
The conclusion presents a YOLO Tiny v4 and Haar-Cascade hybrid for BRTA-compliant Bengali license plates, supported by dynamic retraining and augmentation. Reported results indicate strong recognition performance and practical relevance for resource-constrained real-time deployment.
- 94% precision, 93% recall, and 93% F1 score were achieved for BRTA-compliant Bengali license-plate detection and recognition.
- The system combines YOLO Tiny v4 vehicle detection with Haar-Cascade license-plate localization and a dynamic retraining pipeline.
- Dynamic range enhancement strengthened resilience to lighting variations and occlusions, while the public dataset supports further non-Latin script research.
- The hybrid framework is presented as improving accuracy and computational efficiency for regions such as Bangladesh.