Signature Verification using Siamese Neural Network

This project uses a Siamese Neural Network to verify handwritten signatures and determine whether a signature is Genuine ✅ or Forged ❌. The model compares two input images and predicts similarity based on learned features.

---

Model Used: Siamese Neural Network (SNN)

A Siamese Neural Network is a special type of neural network used to compare the similarity between two inputs. In this project, each "twin" of the SNN is a Convolutional Neural Network (CNN) that extracts features from the two input signature images. The model learns to distinguish genuine and forged signatures based on feature similarity.

---

Why Siamese Neural Network?

• Works well with few samples per class (ideal for signature verification).
• Learns a distance-based similarity function, not just classification.
• Effective in one-shot learning tasks.
• Generalizes well to new, unseen signatures during inference.

---

Features

• Preprocessing of grayscale signature images
• Siamese CNN architecture for feature extraction
• Custom contrastive loss and cosine similarity
• Real-time prediction with image display
• Accuracy improvements via augmentation & tuning

---

How it Works

• Preprocess signature images (resize, normalize).
• Generate pairs of genuine/forged signatures with labels.
• Train a Siamese Network to learn similarity.
• Predict whether a signature matches a reference using a similarity score.

---

Accuracy

CNN vs SNN

Feature	Convolutional Neural Network (CNN)	Siamese Neural Network (SNN)
Input	Single image	Pair of images
Output	Class label	Similarity score (0 = Forged, 1 = Genuine)
Task	Classification	Verification / Similarity Learning
Data Requirement	Many samples per class	Fewer samples, just similar/dissimilar pairs
Use Case	Image classification, object detection	Face/signature verification, one-shot learning

---

Instruction For implementing

Step 1: Downlaod the Kaggle Dataset(CEDAR)

• Here is the link for that=https://www.kaggle.com/datasets/matteocarnebella/cedar-signatures

Step 2: Open Google Colab

• Mount Google Drive to access the project files.

Method: Mount via Python Code

Click the third option (icon in Colab file panel) OR run the code below:

from google.colab import drive
drive.mount('/content/drive')

Step 3: Create folder in Drive

• Create a folder called Data and upload the Kaggle downloaded image
• The path should be

 drive/
    └── MyDrive/
        └── Data/
            └── signatures/
                └── signatures_1/
                    ├── genuine_1_1.png
                    ├── genuine_1_2.png
                    ├── forgeries_1_1.png
                    └── forgeries_1_2.png

Step 4: Run the cells and downlaod the Model

• The code for this is =ipynb/Sure_sign_SNN run the cell and download the model
• The model name is siamese_model2.h5
• Create a folder in the drive called saved_model and the save the model in the folder

/content/drive/MyDrive/Data/saved_model/siamese_model2.h5

Step 5: Upload the file in the google colab(Document)

• A python code is bulid to extract the images and put them in different folders are freehand_signatures and boxed_signatures
• For extracting the signatures
• Click the first option in the colab file panel
• Save the Document file name as signature.png

Step 6: For Verfication and comparing the ouput

• Downlaod the original image from the extracted images
• Same as the 5th step uplaod the original signature signature1.png
• The test images are the freehand_signatures and boxed_signatures

Outcome

The model outputs a similarity score between two signature images. If the score is above a defined threshold (e.g., 0.6), the signatures are classified as Genuine, else Forged.

Genuine Signature

Forged Signature

---

Image Extraction and Verification with the Existing model

• Signature Extraction Code: Extract signatures from scanned documents or image files using preprocessing techniques like resizing, thresholding, and contour detection.
• Original Signature: The reference image used to verify authenticity.
• Extracted Signature (Test Image): The signature image extracted from input, used for comparison.
• Load the Trained Siamese Model: Use the pre-trained Siamese model (siamese_model.h5) by loading it with the custom distance function.
• Verify Signatures: Compare the original and test signature images using the trained model and check the similarity
  • ✅ If similar → "Matching"
  • ❌ If not similar → "Not Matching"
• A simple and effective pipeline for real-time signature verification using a ready-to-use deep learning model.

---

Output of Image extraction and Verification

Orignal Signature

Test Signature Compared with Original image

Test Signature Compared with Original image