Evaluating the Impact of Data Augmentation on MobileNet Performance in Handwritten Signature Identification

Overview This academic project, conducted as a part of a Bachelor's thesis, investigates the effect of affine data augmentation techniques—including rotation, scaling, and translation—on the performance of MobileNetV3-Small for offline handwritten signature identification. The study aims to enhance model generalization and accuracy in biometric verification systems.

Tech Stack and Tools

• Model: MobileNetV3-Small

• Frameworks/Libraries: TensorFlow, Keras, NumPy, scikit-learn, Matplotlib

• Environment: AWS EC2, JupyterLab

• Dataset: BHSig260 Hindi (Kaggle)

Objectives

1. Implement the MobileNetV3-Small architecture for signature identification

2. Develop a data augmentation pipeline using affine transformations

3. Evaluate model performance with and without data augmentation

Methodology

• Applied preprocessing: image resizing to 224×224, normalization

• Used TensorFlow's augmentation layers: RandomRotation, RandomTranslation, RandomZoom

• Trained the model under two conditions—original dataset and augmented dataset

• Evaluated using accuracy, precision, recall, F1 score, and loss

Results

• Accuracy improved from 75.26% (original data) to 85.94% (augmented data)

• Increased F1 score and reduced loss, indicating enhanced generalization

• The model trained on augmented data outperformed the baseline in both validation and test evaluations

Impact The study demonstrates that applying basic, label-preserving data augmentation significantly improves the performance of lightweight models like MobileNetV3-Small in complex tasks such as offline handwritten signature verification. These findings have practical implications for developing efficient and scalable biometric authentication systems.