Rotation-free Online Handwritten Character Recognition Using Linear Recurrent Units

TL;DR

This work tackles rotation-induced degradation in online handwritten character recognition by introducing Sliding Window Path Signature (SW-PS) features coupled with Linear Recurrent Units (LRU) for efficient, parallelizable sequence classification. By capturing local dynamics via SW-PS and modeling long-range dependencies with a complex-valued recurrence, the method achieves robust performance under rotations up to $\pm 180^{\circ}$ and demonstrates rapid convergence. On CASIA-OLHWDB1.1 subsets for digits, English uppercase letters, and Chinese radicals, ensemble SW-PS+LRU achieves accuracies of $99.62\%$, $96.67\%$, and $94.33\%$, respectively, outperforming CNN-based networks and other sequence models in both speed and accuracy. The approach is validated with publicly available data and code, underscoring its practical impact for rotation-free online handwriting recognition and scalable, long-sequence modeling.

Abstract

Online handwritten character recognition leverages stroke order and dynamic features, which generally provide higher accuracy and robustness compared with offline recognition. However, in practical applications, rotational deformations can disrupt the spatial layout of strokes, substantially reducing recognition accuracy. Extracting rotation-invariant features therefore remains a challenging open problem. In this work, we employ the Sliding Window Path Signature (SW-PS) to capture local structural features of characters, and introduce the lightweight Linear Recurrent Units (LRU) as the classifier. The LRU combine the fast incremental processing capability of recurrent neural networks (RNN) with the efficient parallel training of state space models (SSM), while reliably modelling dynamic stroke characteristics. We conducted recognition experiments with random rotation angle up to $\pm 180^{\circ}$ on three subsets of the CASIA-OLHWDB1.1 dataset: digits, English upper letters, and Chinese radicals. The accuracies achieved after ensemble learning were $99.62\%$, $96.67\%$, and $94.33\%$, respectively. Experimental results demonstrate that the proposed SW-PS+LRU framework consistently surpasses competing models in both convergence speed and test accuracy.

Figures (8)

• Figure 1: Data preprocessing workflow.
• Figure 2: Effect of Hanging Normalization in SC mode. The left side is "before", with the true label and rotation angle in the upper left corner. The right side is "after". The start red point is at the top and the center yellow point is at the bottom. Both are located on the same vertical line.
• Figure 3: Illustration of the step-2 truncated signature for a planar path (in red). The displacements $\Delta X^1$ and $\Delta X^2$ are the first level iterated integrals. The signed area $A_+-A_-$ is a linear combination of second iterated integrals based on Green's theorem.
• Figure 4: The design of LRU Block. The Batch Norm mitigates covariate bias. The LRU Layer performs complex-valued linear recursion. Gaussian Error Linear Units (GELU) and Gated Linear Units (GLU) provide nonlinearities through smooth activation and learnable gating respectively. Dropout is applied between activations as a stochastic regularizer to mitigate overfitting and enhance generalization.
• Figure 5: Overall Architecture of the Model.