Register Number Extraction

Live webpage :

https://answer-sheet-proceappr-fjcehvmxe5bxopwvfhnkuv.streamlit.app/

run the script.py with this command :

python script.py --image "rs_test3.png" --yolo_weights "weights.pt" --crnn_model "best_crnn_model(git).pth"

To test the project, follow this link :

https://colab.research.google.com/drive/1TRqyKe7JmP0dOh4_ijVkAqoEycNRgFWP?usp=sharing

1. Tesseract OCR Experiments

This section details the results obtained using Tesseract OCR for register number extraction.

Initial Accuracy: 58.88% (on 1000 iterations)

Accuracy Improvements:

• Increased Iterations (2000): Accuracy improved to 64.11%.
• Increased Iterations (5000) and Regex Application: Accuracy further improved to 78.88% (regex was applied to the test predictions).

Important Notes on Tesseract Experiments:

• Accuracy Variability: Running the training multiple times resulted in variable accuracy, with a fluctuation of approximately (+-) 1%.
• Overfitting: Increasing the number of iterations beyond 5000 did not lead to further positive improvements in accuracy, suggesting potential overfitting of the Tesseract model to the training data.

2. Custom Model Experiments

This section outlines the results achieved with custom-trained models for register number extraction.

Initial Custom Model Accuracy (without Data Augmentation): 95.14%

Impact of Data Augmentation and Increased Epochs:

• Applying data augmentation and increasing the number of training epochs to 5 resulted in an accuracy of 93%.

Detailed Experiment with "deep CRNN TEMP (with data augmentation).ipynb":

• 70 Epochs: Achieved an accuracy of approximately 91%, with mediocre prediction quality.
• 75 Epochs: Increased the accuracy to around 92.9%, and the prediction quality was observed to be significantly better.
• 80 Epochs: Resulted in a higher accuracy of 94%, but the prediction quality deteriorated, indicating potential overfitting.

Note on Model Saving and Loading:

• Previously, there were issues encountered while saving and loading the trained models. Efforts should be made to ensure the complete and reliable saving of the best-performing model.

3. Best Models to Consider

Based on the experiments, the following model notebooks appear to be the most promising and should be considered for further development or deployment:

• CRNN with just dropout.ipynb
• deep CRNN TEMP (with data augmentation).ipynb

Setting up the project on your PC :

What You’ll Need

• Computer: Windows, Linux, or macOS.
• Python: Version 3.8 to 3.13 (we recommend 3.10, but newer ones like 3.12 work too).
• Images: PNG files with 12-digit numbers (e.g., 21222xxxxxx.png).

• NOTE : All the Register Numbers are under the folder data/Register Numbers, if you face any file not found errors pls verify your path!

• GPU (Optional): An NVIDIA graphics card with CUDA 12.1 for faster performance (e.g., RTX 4060).

1. Get the Code

First, download this project to your computer:

1. Open your terminal (Command Prompt on Windows, Terminal on Mac/Linux).

2. Type this command and press Enter:

   git clone https://github.com/SANTHAN-KUMAR/Register-Number-Extraction
   

2. Move into the project folder:

cd Register-Number-Extraction

3. Set Up a Space for the Code

We’ll create a “virtual environment” to keep things organized:

1. Run this command:

python -m venv env

2. Activate it :

env\Scripts\activate

4. Install the Tools

This project needs some Python libraries. Pick the option below based on your setup:

Option A: Regular Computer (CPU)

If you don’t have an NVIDIA GPU or just want to keep it simple:

1. Run :

pip install -r requirements.txt

2. Check if it worked :

python -c "import torch; print('PyTorch version:', torch.__version__); print('GPU available?', torch.cuda.is_available())"

It should output something like this :

PyTorch version: 2.5.1
GPU available? False

Option B: NVIDIA GPU (Faster Setup)

If you have an NVIDIA GPU with CUDA 12.1 (like the RTX Series):

1. Check your CUDA version:

nvcc --version

Look for “release 12.1” (e.g., Cuda compilation tools, release 12.1, V12.1.105). If it’s different (e.g., 11.8), see the “Other GPUs” note below.

2. Install the GPU version:

pip install torch==2.5.1+cu121 torchvision==0.20.1+cu121 -f https://download.pytorch.org/whl/torch_stable.html
pip install numpy>=1.26.0 pillow>=10.1.0

3. Check if it worked:

python -c "import torch; print('PyTorch version:', torch.__version__); print('GPU available?', torch.cuda.is_available())"

You should see :

PyTorch version: 2.5.1+cu121
GPU available? True

Note for Other GPUs: If your CUDA version isn’t 12.1 (e.g., 11.8), use this instead:

pip install torch==2.5.1+cu118 torchvision==0.20.1+cu118 -f https://download.pytorch.org/whl/torch_stable.html
pip install numpy>=1.26.0 pillow>=10.1.0

Find your version at PyTorch’s site.

4. Run the Project :

1.Start Jupyter Notebook:

jupyter notebook

A browser window will open.

Open the .ipynb file (deep CRNN TEMP (with data augmentation)-updated.ipynb).

Click “Run All” in the menu, or run each cell one-by-one:

Loads your images.

Trains the model (takes time, especially on CPU).

Tests it and predicts numbers from images.

The code picks CPU or GPU automatically, so no changes needed!

What’s in the Project:

• Cells 1-2: Load and prepare your images.

• Cell 3: Defines the CRNN (the brain of the project).

• Cell 4: Trains it and saves the best version to saved_models/.

• Cell 5: Tests accuracy on some images.

• Cell 6: Predicts a number from one image (e.g., test_reg.jpeg).

To take part of this project, fork this repository and give a pull request with your changes! Have a look at the issues mentioned, it is a great way to contribute to this project by solving them!