TLDR: Matheus Ramos Parracho’s research explores improving signature forgery detection by enhancing cross-dataset generalization. The study uses Siamese neural networks with ResNet-34 backbones, trained on raw images and a novel ‘shell-based’ pre-processing method. While raw image models achieve high performance on familiar datasets, the shell-based approach, which extracts 1D contours, demonstrates more consistent performance across diverse datasets like CEDAR, ICDAR, and GPDS Synthetic, reducing dataset-specific biases and improving overall robustness.
Automated signature verification is a crucial technology used in many sectors, from banking to legal documents, to confirm identity. However, the wide variety of handwriting styles and increasingly sophisticated forgery techniques make it a complex challenge to accurately distinguish between authentic and fake signatures. A significant hurdle in current offline signature verification systems is their limited ability to perform well across different datasets, meaning a model trained on one set of signatures often struggles when presented with signatures from another.
Matheus Ramos Parracho, in his 2025 research paper, “Signature Forgery Detection: Improving Cross-Dataset Generalization,” addresses this critical limitation. The study aims to develop robust models capable of handling signatures from diverse datasets with consistent performance, thereby enhancing their reliability against variations in handwriting and forgery methods. This research, conducted at the Universidade Federal do Rio de Janeiro, explores advanced feature learning techniques using neural networks with a Siamese architecture and custom layers.
The work is divided into two main parts. The first part focuses on enhancing model generalization by investigating and developing signature verification techniques that use raw image data. The goal here is to achieve stable and satisfactory performance when moving across different datasets with varied handwriting styles.
The methodology involved using three widely recognized datasets: CEDAR, ICDAR, and GPDS Synthetic. These datasets offer a diverse range of genuine and forged signatures. A key aspect of the data preparation was a unique pairing strategy. Beyond comparing genuine signatures from the same writer (labeled as authentic) and genuine-forged pairs from the same writer (labeled as forged), the researchers also included genuine signatures from *different* writers, labeling these as ‘forged’ pairs. This innovative approach was found to be essential for improving the model’s ability to generalize across different datasets, teaching it to distinguish between different identities rather than just memorizing specific patterns.
The models developed in this first part utilized a Siamese Neural Network with a ResNet-34 backbone, trained with either Contrastive Loss or Triplet Loss. Contrastive Loss aims to pull similar embeddings closer and push dissimilar ones apart, while Triplet Loss enforces a stricter ordering, ensuring that an anchor signature is closer to a genuine counterpart than to a forged one by a certain margin. The results showed that models trained on raw images could generalize well, with Triplet Loss, despite being more challenging to train, demonstrating superior cross-dataset generalization capabilities.
The second part of the research introduces a novel approach: designing a shell-based pre-processing pipeline. The primary objective here was to reduce the complexity and dimensionality of the input data while retaining crucial information about the signature’s structure. This pipeline transforms the 2D signature image into a set of 1D functions, essentially extracting successive horizontal contours, or “shells,” from the ink trace.
The shell extraction process involves several steps: binarizing the raw grayscale image, extracting a bounding box, cropping, and resizing to a standard 512×512 resolution. Then, a custom algorithm extracts superior, inferior, and residual shells, which are essentially 1D vectors representing the vertical profile of the signature. Additionally, auxiliary features like approximated pressure (based on grayscale intensity) and thickness were considered, though thickness was implicitly captured by the shell differences.
For this shell-based data, a simpler model architecture was designed, including a specialized “PressureShell Layer” (PS Layer) to integrate shell and pressure signals, followed by convolutional blocks and fully connected layers. An adapted one-dimensional ResNet-34 backbone was used for training. Interestingly, initial attempts to normalize the final embeddings onto a hypersphere led to training instability; removing this step significantly improved stability and convergence.
The results from the shell-based approach revealed a fascinating outcome: a “performance equalization” across datasets. While raw image models often achieved higher peak performance on datasets they were trained on (like CEDAR and ICDAR), their performance could drop significantly on a more challenging, unseen dataset like GPDS. The shell-preprocessed models, however, showed more balanced results across all datasets, mitigating large discrepancies. This suggests that while some fine discriminative details might be lost in the preprocessing, the shell representation helps reduce dataset-specific biases, leading to more consistent generalization. The code for this research is open-source and available at https://github.com/mathparracho/general signature forgery detector.
Also Read:
- Decoding the Past: ClapperText and Low-Resource Text Recognition
- Securing AI’s Foundation: A New Method for Verifying Dataset Ownership in Language Models
In conclusion, this study successfully implemented a comprehensive training pipeline for Siamese networks, emphasizing cross-dataset evaluation and demonstrating the effectiveness of ResNet-based architectures for signature verification. Furthermore, the introduction of a shell-based preprocessing method offers a promising alternative for creating more robust and generalizable signature verification systems, even if it means a slight trade-off in peak performance for improved consistency across diverse data. Future work could explore more advanced loss functions, smarter sampling techniques, and meta-learning to further enhance these systems.
