Handwritten and Photocopied Signature Datasets for Fraud Detection (HPS, HBS, HMS)
Description
This repository contains three signature datasets collected for the study "Authentic Signature Recognition and Fraud Detection Using Transfer Learning" by Sharifur Rahman. The archive includes: (1) the Handwritten–Photocopy Signature (HPS) dataset, comprising 2,705 handwritten and 2,705 photocopied signature images acquired using various scanning and printing devices, developed to distinguish original handwritten signatures from photocopied reproductions; (2) the Handwritten Bangla Signature (HBS) dataset, containing 2,480 signature samples from 155 writers, with eight genuine and eight forged signatures per writer; and (3) the Handwritten Mixed Signature (HMS) dataset, containing 6,320 samples from 395 writers, including Bangla, English, and graphical signatures, likewise with eight genuine and eight forged signatures per writer. These datasets were used to train and evaluate transfer learning-based feature extractors (VGG16, ResNet50, Inception-v3, Xception) combined with classical machine learning classifiers (Random Forest, Linear SVM, RBF SVM, Sigmoid SVM) for signature verification and forgery detection. The data may support research in offline signature verification, document forgery detection, biometric authentication, and forensic document analysis.
Files
Steps to reproduce
Steps to Reproduce Authentic Signature Recognition and Fraud Detection Using Transfer Learning 1. Download and extract the data Download and extract the archive (Signature_Datasets.zip), which contains three sub-folders: HPS, HBS, and HMS, corresponding to the three datasets described in the dataset overview. 2. Preprocessing Apply histogram equalization or contrast stretching to enhance stroke visibility, followed by noise reduction to improve image quality. Resize all signature images to 224 x 224 pixels to ensure a consistent input size for the CNN feature extractors. 3. Data augmentation (training set only) Apply geometric transformations, including flipping, rotation, translation, cropping, and scaling, along with variations in brightness, color, and image noise, to increase the quantity and diversity of the training data. 4. Label encoding Convert categorical class labels (genuine/forged, or handwritten/photocopied for the HPS dataset) into numerical form using one hot encoding. 5. Feature extraction Pass the preprocessed images through four pre trained CNN architectures, VGG16, Inception v3, ResNet50, and Xception, each fine tuned separately using the Adam and RMSprop optimizers, to obtain feature vectors (512 dimensional for VGG16; 2048 dimensional for the others). 6. Classification Train and evaluate Random Forest, Linear SVM, RBF SVM, and Sigmoid SVM classifiers on the extracted feature representations. For the HPS dataset specifically, also train a custom CNN and a VGG16 model in an end to end configuration for handwritten versus photocopy classification. 7. Train/test split For the HPS dataset, use a device aware, cross device evaluation protocol: train on signatures from one group of acquisition devices and test on signatures from different, previously unseen devices. For the HBS and HMS datasets, use a writer blind split so that signatures from the same writer do not appear in both the training and testing sets. 8. Evaluation Compute Accuracy, Precision, Recall, F1 Score, Average Error Rate (AER), and Equal Error Rate (EER) for each classifier and feature extractor combination, following the formulas described in the accompanying manuscript (Section 3.5). 9. Reference For full experimental details, hyperparameters, and result tables, refer to the published article: Authentic Signature Recognition and Fraud Detection Using Transfer Learning.
Institutions
- Noakhali Science and Technology UniversityChittagong, Chittagong
- Green University of BangladeshDhaka Division, Dhaka