Introduction
Signature verification is a critical component of document authentication and retrieval. However, real-world scenarios often present "partial" signatures—initials, clipped scans, or signatures obscured by machine-printed text. Traditional holistic methods, which view the signature as a single global shape, fail when parts are missing. Conversely, point-level matching is computationally expensive and scales poorly. This paper proposes a robust, part-based framework that treats signatures as "documents" composed of "salient regions" (topics), leveraging supervised Latent Dirichlet Allocation (sLDA) and Hierarchical Dirichlet Processes (HDP) for efficient, scalable matching.
Methodology
Feature Extraction: Modified Shape Context
The foundation of the approach is a modified version of Shape Context features. Standard shape context describes the distribution of points around a contour point relative to the local signature scale. To handle partial signatures, the authors normalize distances across the entire dataset rather than per image. This global normalization ensures that a specific stroke (e.g., a loop) retains the same feature representation whether it appears in a full signature or an isolated fragment.
The "Signature-as-Document" Model
The core innovation is the application of topic modeling to handwriting. By treating a signature as a document, the authors define a hierarchy where visual words (discretized shape features) form salient regions (topics), and a signature is a distribution over these regions. The author's identity serves as the label in a supervised learning framework.
To handle the unknown number of salient regions across a diverse dataset, the authors employ Hierarchical Dirichlet Processes (HDP), allowing the model to automatically determine the optimal complexity for describing an author's style.
Experimental Results
The model was evaluated on the clean Tobacco dataset (DS-I) and the large-scale, "in-the-wild" UMD dataset (DS-II), which features real-world degradations like low resolution and overlapping text.
On the DS-I partial dataset, the proposed method achieved 87.8% Top-1 accuracy, vastly outperforming Locality Sensitive Hashing (LSH) at 17.5%. Even on the highly degraded DS-II dataset, the model achieved 37.8% accuracy, outperforming the LSH baseline by over 23%. These results demonstrate that the model is not only robust to missing data but also effective as a general-purpose signature matcher, hitting 92.1% accuracy on full, clean signatures.
Conclusion
By reframing signature matching as a topic modeling problem, this framework provides a scalable solution for identifying authors from partial data. The combination of global feature normalization and hierarchical graphical models allows the system to capture the "visual vocabulary" of an author's handwriting, moving beyond rigid global shape matching toward a more flexible, part-based understanding of identity. Future work will focus on improving performance in extremely noisy environments where signatures are heavily occluded by background document elements.
Summary updated