Handwriting Signature Writer Identification using CNN

This project implements a Convolutional Neural Network (CNN) model to identify the writer of a given handwritten signature. The model is trained on a dataset of genuine and forged signatures, and it aims to classify a new signature sample to its respective writer.

Developed in Google Colab using TensorFlow and Keras.

---

Table of Contents

• Project Overview
• Dataset
• Methodology
  • Data Preprocessing
  • Model Architecture
  • Training
• Results
• How to Use
  • Prerequisites
  • Setup
  • Running the Notebook
• Future Work
• Author
• Acknowledgments

---

Project Overview

The goal is to build a system that identifies the author of a handwritten signature. This has applications in forensics and document verification. A CNN is used due to its effectiveness in image recognition tasks.

---

Dataset

The dataset contains handwritten signatures from multiple writers, including both genuine (full_org) and forged (full_forg) samples.

• Source: [Add dataset name/link or brief explanation if private]
• Structure:
  • Example filenames: original_58_1.png, forgeries_1_10.png
• Image Format: PNG, TIFF, or JPG

Note on Privacy: All data has been anonymized and used solely for research purposes. Ensure proper consent when using real signature data.

---

Methodology

Data Preprocessing

• Loading: Signature paths collected from Google Drive
• Labeling: Writer IDs extracted from filenames
• Encoding: String IDs converted to integers
• Splitting: 80/20 train-validation split (with stratification)
• Image Processing:
  • Resize to 64x256
  • Grayscale (1 channel)
  • Normalize pixel values [0, 1]

TensorFlow Dataset Pipeline is used for efficient loading, batching, and prefetching.

---

Model Architecture

CNN model structure:

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #
=================================================================
 conv2d (Conv2D)             (None, 62, 254, 32)       320
 max_pooling2d (MaxPooling2D) (None, 31, 127, 32)      0
 conv2d_1 (Conv2D)           (None, 29, 125, 64)       18496
 max_pooling2d_1 (MaxPooling2D) (None, 14, 62, 64)     0
 conv2d_2 (Conv2D)           (None, 12, 60, 128)       73856
 max_pooling2d_2 (MaxPooling2D) (None, 6, 30, 128)     0
 flatten (Flatten)           (None, 23040)             0
 dense (Dense)               (None, 128)               2949248
 dropout (Dropout)           (None, 128)               0
 dense_1 (Dense)             (None, 55)                7095
=================================================================
Total params: 3,046,015
Trainable params: 3,046,015
Non-trainable params: 0
_________________________________________________________________

---

Training

• Optimizer: Adam
• Loss: SparseCategoricalCrossentropy(from_logits=True)
• Metrics: Accuracy
• Epochs: 30 (or adjust as needed)

Training history includes plots for accuracy and loss across epochs.

---

Results

On validation set:

• Validation Accuracy: 95.83%
• Validation Loss: 0.25

Training and Validation Accuracy:

Training and Validation Loss:

Example Test (Writer ID: 58):

• Accuracy: 95.83% (23/24 correct)

How to Use

Prerequisites

• Google Colab or local Python environment
• Python Libraries:
  • TensorFlow (2.x)
  • NumPy
  • Matplotlib
  • scikit-learn
  • Pillow

Setup

1. Clone the repo (optional):

    git clone https://github.com/TranHuuDat2004/handwriting-signature-recognition.git
    cd handwriting-signature-recognition
   

2. Upload Notebook to Colab

3. Prepare Dataset: - Upload dataset to Google Drive - Update BASE_DATA_DIR in the notebook:

        BASE_DATA_DIR = '/content/drive/MyDrive/YOUR_PATH_TO/SIGNATURES'
   

---

Running the Notebook

1. Mount Google Drive
2. Run cells sequentially: - Collect paths and labels - Preprocess images - Create Dataset objects - Build and train the model - Evaluate and save model - Predict single or batch samples - Evaluate on a specific writer ID

---

Future Work

• Try ResNet, VGG, or MobileNet
• Improve data augmentation
• Explore Siamese Networks for signature verification
• Build a web interface (e.g., with Streamlit)
• Train on larger datasets

---

Author

Tran Huu Dat
GitHub: @TranHuuDat2004

---