This project implements a deep learning-based signature verification system using a Siamese Neural Network. The model learns to compare two signature images and determine whether they are from the same person (genuine) or not (forged).
This system is designed for offline signature verification using a Siamese neural network architecture. It employs Convolutional Neural Networks (CNNs) to extract features from input images and compares them using Euclidean distance to determine similarity.
- Base CNN Encoder: Shared network to extract features from both input images
- Distance Function: Euclidean distance computed between feature vectors
- Loss Function: Mean Squared Error (MSE)
- Optimizer: Adam
The model is trained to reduce the distance between similar signatures and increase the distance for dissimilar ones.
The training data should be provided in a CSV file (train_data.csv) with the following format:
image1,image2,label 001/001_01.PNG,001/001_02.PNG,1 001/001_01.PNG,001_forg/001_05.PNG,0
1: Same person (genuine signature pair) 0: Different persons (forged pair)
Install the required dependencies using pip:
pip install tensorflow numpy pandas opencv-python
Ensure the train_data.csv and signature image folders are correctly set up. Then, run the training script:
python signature_verification.py
This will:
- Load and preprocess the image pairs
- Train the Siamese neural network
- Save the trained model as siamese_model.h5
You can verify if a test signature matches a reference signature using the verify_signature() function:
result = verify_signature(model, "test_signature.png", "reference_signature.png", threshold=0.5) print(result)
Output:
- "Test Signature is authentic."
- "Test Signature is a forgery."
To save the trained model:
model.save('siamese_model.h5')
To load the saved model:
from tensorflow.keras.models import load_model model = load_model('siamese_model.h5', custom_objects={'euclidean_distance': euclidean_distance})
- Siamese network architecture for one-shot learning
- Simple image preprocessing (grayscale, resize, normalize)
- Flexible CSV-based pairing system
- Easily extensible and production-ready