Vehicle Registration Plate Detection using Detectron2

This project implements a vehicle registration plate detection system using Detectron2, a powerful object detection and segmentation framework built on PyTorch. The project involves training a model on a custom dataset to identify and localize registration plates in images.

Installation

1. Install Detectron2 and other dependencies:

!pip install pycocotools

2. Set up the Detectron2 environment:

import detectron2
from detectron2.utils.logger import setup_logger
setup_logger()

3. Import necessary libraries:

import numpy as np
import cv2, random, os
import matplotlib.pyplot as plt
from detectron2 import model_zoo
from detectron2.engine import DefaultTrainer, DefaultPredictor
from detectron2.config import get_cfg
from detectron2.utils.visualizer import Visualizer
from detectron2.structures import BoxMode
from detectron2.data import DatasetCatalog, MetadataCatalog
from detectron2.evaluation import COCOEvaluator, inference_on_dataset
from detectron2.data import build_detection_test_loader

---

Dataset Preparation

1. Directory Structure

   • Ensure the dataset is organized as follows:

     /Dataset
     |-- train
     |   |-- Vehicle registration plate
     |       |-- Images
     |       |-- Label
     |-- validation
         |-- Vehicle registration plate
             |-- Images
             |-- Label
     

2. Annotations

   • Annotations are stored in .txt files, with bounding box coordinates for each image.
   • Example of an annotation line:

     class_id x_center y_center width height xmin ymin xmax ymax
     

3. Dataset Registration

   • Register the datasets for training and validation using the following code:

DatasetCatalog.register(name='train_regn_plate', func=lambda: get_regn_plate_dicts(data_root, train_txt))
DatasetCatalog.register(name='val_regn_plate', func=lambda: get_regn_plate_dicts(data_root, val_txt))
MetadataCatalog.get('train_regn_plate').set(thing_classes=['regn_plate'])
MetadataCatalog.get('val_regn_plate').set(thing_classes=['regn_plate'])

---

Code Overview

Dataset Registration

The function get_regn_plate_dicts processes the annotations and image metadata into a Detectron2-compatible format. It:

• Reads image dimensions and paths.
• Parses bounding box annotations.
• Outputs a list of dictionaries containing image and annotation data.

Visualization

Visualize random samples from the dataset to verify correct registration:

for d in random.sample(val_data_dict, 3):
    img = cv2.imread(d["file_name"])
    visualizer = Visualizer(img[:, :, ::-1], metadata=val_metadata, scale=0.5)
    vis = visualizer.draw_dataset_dict(d)
    plt.figure(figsize=(12, 12))
    plt.imshow(vis.get_image())
    plt.show()

Training

1. Configure the model using a pre-trained RetinaNet from the Detectron2 model zoo.
2. Customize solver parameters and dataset settings:

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file('COCO-Detection/retinanet_R_50_FPN_3x.yaml'))
cfg.DATASETS.TRAIN = ('train_regn_plate',)
cfg.DATASETS.TEST = ()
cfg.DATALOADER.NUM_WORKERS = 4
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url('COCO-Detection/retinanet_R_50_FPN_3x.yaml')
cfg.SOLVER.IMS_PER_BATCH = 8
cfg.SOLVER.BASE_LR = 0.0001
cfg.SOLVER.MAX_ITER = max_iter
cfg.MODEL.RETINANET.NUM_CLASSES = 1
cfg.OUTPUT_DIR = 'outputs'

3. Train the model:

trainer = DefaultTrainer(cfg)
trainer.resume_or_load(resume=False)
trainer.train()

Inference

Perform predictions on validation data and visualize the results:

predictor = DefaultPredictor(cfg)
for d in random.sample(val_data_dict, 5):
    im = cv2.imread(d["file_name"])
    outputs = predictor(im)
    v = Visualizer(im[:, :, ::-1], metadata=val_metadata, scale=0.8)
    v = v.draw_instance_predictions(outputs['instances'].to('cpu'))
    plt.figure(figsize=(12, 12))
    plt.imshow(v.get_image())
    plt.show()

---

Evaluation

1. Evaluate the model on the validation dataset using COCO metrics:

evaluator = COCOEvaluator(dataset_name='val_regn_plate', output_dir=eval_dir, distributed=False)
val_loader = build_detection_test_loader(cfg, 'val_regn_plate')
inference_on_dataset(trainer.model, val_loader, evaluator)

Acknowledgments

• Detectron2: A modular and scalable object detection library.
• Dataset: Custom dataset for vehicle registration plate detection.