fastplateocr-py 0.1.2

Accurate and Efficient General OCR System for License Plates

FastPlateOCR

Visual samples of challenging real-world license plates (motion blur, diverse layouts, low light) that FastPlateOCR is built to handle.

🚀 FastPlateOCR is an accurate, extremely fast, and flexible End-to-End License Plate Recognition library.

Unlike traditional ALPR systems that rely on heavy architectures, FastPlateOCR introduces cutting-edge structural improvements designed specifically for real-time edge deployment. Our framework achieves ultra-low inference latency without sacrificing accuracy on blurry or degraded license plates through two major architectural optimizations.

🧩 FastPlateOCR Pipeline

The framework is structured as a highly optimized two-stage sequential pipeline:

➔ Improved YOLO26n ➔ ➔ Improved SVTRv2-Tiny ➔ 59P289136

Overview of the proposed highly optimized two-stage ALPR pipeline.

1. Improved YOLO26n for Fast Detection

We replaced the Original sequential Bottleneck blocks in the YOLO26 neck with our novel 5x5 RepMixer blocks. By leveraging structural re-parameterization, the model trains with a rich multi-branch topology (capturing complex spatial contexts) but mathematically fuses into a single, highly efficient 5x5 convolution during inference. This completely eliminates memory fragmentation and intermediate read/write operations, drastically reducing Memory Access Cost (MAC).

Original: Sequential Bottleneck	Proposed: RepMixer (Inference Fused)

2. Improved SVTRv2-Tiny for Lightning-Fast Recognition

To make the SVTRv2 OCR model viable for strict real-time constraints, we applied two key modifications:

• RepMixer-enhanced Feature Extraction: We replaced Original standard convolutions in the early stages with RepMixer blocks. A single fused 5x5 kernel efficiently captures continuous morphological strokes (like loops and lines in characters) without the latency penalty of stacked 3x3 convolutions.

Original: Standard Convolution	Proposed: RepMixer

• Efficient RCTC Decoder: We entirely discarded the Original heavy attention-based RCTC Decoder. Since license plates have a rigid, horizontally aligned structure, we replaced 2D attention with a simple Height-wise Average Pooling operation. This elegantly compresses the 2D features into a 1D sequence, completely bypassing expensive matrix multiplications.

Original: Heavy RCTC Decoder	Proposed: Efficient RCTC Decoder

By integrating these specialized components, FastPlateOCR delivers unmatched production-ready performance, processing frames at blazing speeds!

---

🛠 Installation

pip install fastplateocr-py

(Note: To use the auto-download feature for pre-trained weights, please ensure huggingface_hub is installed).

⚡ Quick Start

FastPlateOCR automatically downloads the best pre-trained models from our HuggingFace repository the first time you run it. You don't need to manually configure any paths!

1. End-to-End Recognition (Detect & Read)

import cv2
from fastplateocr import FastPlateOCR

# Initialize (auto-downloads weights if not found)
model = FastPlateOCR()

# Read the plate
results = model.read('car_image.jpg')

for res in results:
    print(f"Plate Text: {res['text']} | Confidence: {res['score']:.4f}")
    print(f"Bounding Box: {res['box']}")

2. Flexible API: Detect Only

If you only need to locate the license plates without reading the text:

# Disable the recognition model to save memory
model = FastPlateOCR(use_rec=False)
boxes = model.detect('car_image.jpg')
print("Detected boxes:", boxes)

3. Flexible API: Recognize Only

If you already have a cropped image of a license plate and just want to read the characters:

# Disable the detection model
model = FastPlateOCR(use_det=False)

crop_img = cv2.imread('cropped_plate.jpg')
text, score = model.recognize(crop_img)
print(f"Text: {text} (Score: {score})")

4. Using Custom Local Weights

If you have fine-tuned your own models or downloaded the weights locally, you can easily load them:

model = FastPlateOCR(
    det_model_path="/path/to/your/yolo.pt",
    rec_model_path="/path/to/your/svtr.pth"
)

🏋️ Training & Evaluation

FastPlateOCR provides a complete suite of scripts in the tools/ directory for dataset preparation, training, evaluation, and inference.

0. Model Weights Preparation

Before training or evaluation, download the official pre-trained models from our HuggingFace Repository and place them in the following structure:

FastPlateOCR/
├── pretrained_models/
│   ├── yolo26n_rep_mixer/
│   │   └── best.pt
│   └── svtrv2_tiny_efficient_rctc/
│       └── best.pth

You can download them manually or use wget:

wget -O pretrained_models/det/yolo26n_rep_mixer/best.pt https://huggingface.co/anhone3/FastPlateOCR/resolve/main/yolo26n_rep_mixer/best.pt
wget -O pretrained_models/rec/svtrv2_tiny_efficient_rctc/best.pth https://huggingface.co/anhone3/FastPlateOCR/resolve/main/svtrv2_tiny_efficient_rctc/best.pth

1. Data Preparation (Create LMDB)

Because our LMDB script uses hardcoded paths for simplicity, please open tools/create_lmdb_dataset.py and modify the data_dir variable in the __main__ block to match your dataset path before running:

if __name__ == '__main__':
    data_dir = './dataset/rec' # Set your dataset directory

    label_file_list = [
        os.path.join(data_dir, 'train_labels.txt'),
        os.path.join(data_dir, 'val_labels.txt'),
        os.path.join(data_dir, 'test_labels.txt')
    ]

After modifying the paths, generate the LMDB:

python tools/create_lmdb_dataset.py

2. Training (Det & Rec)

Before training, you must configure the dataset paths, batch sizes, and learning rates in the respective .yml files.

For Detection (configs/det/yolo26/yolo26n_rep_mixer.yml):

Train:
  data: './dataset/det/data.yaml' # Point this to your YOLO data.yaml
  epochs: 50
  batch: 256

For Recognition (configs/rec/svtrv2/svtrv2_tiny_efficient_rctc.yml):

Train:
  dataset:
    name: RatioDataSetTVResize
    data_dir_list: ['./dataset/rec/lmdb_data/train']

Eval:
  dataset:
    name: RatioDataSetTVResize
    data_dir_list: ['./dataset/rec/lmdb_data/val']

Once configured, start training:

[!TIP] Pre-trained Models (Fine-tuning) By default, the training process will load pre-trained weights to speed up convergence. You can change this path or leave it empty (to train from scratch) by editing the Global.pretrained_model field inside the .yml config files:

Global:
  pretrained_model: './pretrained_models/det/yolo26n_rep_mixer/best.pt'

# Train Detection Model (YOLO26)
python tools/train_det.py -c configs/det/yolo26/yolo26n_rep_mixer.yml

# Train Recognition Model (SVTRv2)
python tools/train_rec.py -c configs/rec/svtrv2/svtrv2_tiny_efficient_rctc.yml

3. Evaluation (Validation)

Evaluate your trained checkpoints on the validation set using the config files:

# Evaluate Detection (using default pretrained_model path from config)
python tools/eval_det.py -c configs/det/yolo26/yolo26n_rep_mixer.yml

# Evaluate Recognition (using default pretrained_model path from config)
python tools/eval_rec.py -c configs/rec/svtrv2/svtrv2_tiny_efficient_rctc.yml

(Optional) You can also override the model path on-the-fly using the -m flag to evaluate your own newly trained weights:

# Evaluate Detection with a custom trained checkpoint
python tools/eval_det.py -c configs/det/yolo26/yolo26n_rep_mixer.yml -m output/det/yolo26n_rep_mixer/train/weights/best.pt

# Evaluate Recognition with a custom trained checkpoint
python tools/eval_rec.py -c configs/rec/svtrv2/svtrv2_tiny_efficient_rctc.yml -m output/rec/svtrv2_tiny_efficient_rctc/train/best.pth

4. Batch Inference

Test your checkpoints directly on directories of images (supports --save_log to save predictions):

# Infer Detection
python tools/infer_det.py -m pretrained_models/det/yolo26n_rep_mixer/best.pt -d dataset/det/test/images --save_log

# Infer Recognition
python tools/infer_rec.py -m pretrained_models/rec/svtrv2_tiny_efficient_rctc/best.pth -d dataset/rec/test --save_log

🤝 Acknowledgements

• OpenOCR: FastPlateOCR is built upon the robust foundation of OpenOCR.
• YOLO26 & SVTRv2: This work heavily leverages the architectural innovations from YOLO26 for high-speed object detection and SVTRv2 for accurate text recognition.
  • Read the YOLO26 Paper
  • Read the SVTRv2 Paper

📧 Contact

For any questions or issues, please open an issue or contact: [e-mail].