[Paper Review] Offline Writer Identification based on the Path Signature Feature

Songxuan Lai, Lianwen Jin|arXiv (Cornell University)|May 3, 2019

Handwritten Text Recognition Techniques35 references4 citations

TL;DR

This paper proposes a novel offline writer identification method using path signature (PS) features extracted from handwriting contours, specifically leveraging log path signatures (LPS) for compact, discriminative representation. The approach achieves state-of-the-art performance on IAM, Firemaker, CVL, and ICDAR2013 datasets, with top-1 accuracies of 97% on IAM, 99.27% on CVL, and 96.6% on ICDAR2013, demonstrating high efficiency and robustness even with limited ink or single-line queries.

ABSTRACT

In this paper, we propose a novel set of features for offline writer identification based on the path signature approach, which provides a principled way to express information contained in a path. By extracting local pathlets from handwriting contours, the path signature can also characterize the offline handwriting style. A codebook method based on the log path signature---a more compact way to express the path signature---is used in this work and shows competitive results on several benchmark offline writer identification datasets, namely the IAM, Firemaker, CVL and ICDAR2013 writer identification contest dataset.

Motivation & Objective

• To develop a principled, data-efficient method for offline writer identification that captures handwriting style from contour geometry.
• To adapt the path signature framework—originally for time series—to offline handwriting by treating contour fragments as paths.
• To improve performance on benchmark datasets without complex preprocessing or feature fusion.
• To evaluate robustness under low-ink conditions and varying template quality.

Proposed method

• Extract local pathlets from handwriting contours by polygonizing the stroke boundaries.
• Compute the path signature (PS) using iterated integrals of the path, capturing geometric and directional features like curvature and orientation.
• Use the log path signature (LPS) to compress the PS into a more compact, stable representation.
• Construct a codebook from LPS features using vector quantization for efficient matching.
• Apply a nearest-neighbor classifier with LPS-based features for writer identification.
• Optimize hyperparameters (w, m, M) for pathlet length, sampling rate, and codebook size to balance accuracy and speed.

Experimental results

Research questions

• RQ1 Can the path signature framework effectively capture discriminative handwriting style from offline handwriting contours?
• RQ2 How does the LPS-based codebook method compare to existing contour-based and deep learning methods on standard benchmarks?
• RQ3 How robust is the method to variations in ink amount and template quality?
• RQ4 Can the method achieve high accuracy with minimal data, such as single text lines or limited templates?

Key findings

• The proposed LPS-based method achieves a top-1 accuracy of 97% on the IAM dataset, outperforming most prior contour-based methods.
• On the CVL dataset, the method reaches 99.27% top-1 accuracy, matching or exceeding state-of-the-art results without feature fusion.
• With only one text line as query, the method achieves 95.31% top-1 accuracy on CVL, demonstrating strong data efficiency.
• On the ICDAR2013 dataset, the method achieves 96.6% top-1 accuracy, showing strong generalization across diverse handwriting styles.
• The method is computationally efficient, with LPS computation significantly faster than SIFT, and performs well even with low-ink templates when hyperparameters are tuned.
• Performance degrades with insufficient ink in templates, highlighting the need for robustness improvements under low-ink conditions.