Synthetic Handwriting with Bézier Curves

• Synthetic handwriting data generation using Bézier curves is a technique that parameterizes handwriting strokes into scalable, variable-degree curves for realistic data synthesis.
• It employs a robust pipeline including style initialization, stroke decomposition, augmentation, and rendering to ensure diverse and natural handwriting outputs.
• The approach supports advanced applications such as HTR, post-OCR correction, and digital calligraphy by accurately mimicking real handwriting error patterns and visual nuances.

Synthetic handwriting data generation using Bézier curves refers to the algorithmic process of creating high-fidelity, scalable representations of natural handwriting by parameterizing strokes as Bézier curves. This approach enables the synthesis of large and stylistically variable handwriting datasets suited for tasks such as Handwritten Text Recognition (HTR) model training and Post-OCR Correction (POC), with precise control over the generated style and error patterns (Davydkin et al., 2023). Modern frameworks leverage these principles for both vector graphics and rasterized output, with recent models generalizing to variable-degree Bézier primitives and employing deep generative architectures for scalable handwriting synthesis and vectorization (Das et al., 2021).

1. Mathematical Foundations of Bézier-Based Handwriting Synthesis

The mathematical core of synthetic handwriting using Bézier curves is the representation of each handwritten stroke as a parametric curve defined by a set of control points. In the cubic case, every stroke is modeled as:

$$B(t) = (1-t)^3 P_1 + 3(1-t)^2 t\,P_2 + 3(1-t)\,t^2 P_3 + t^3 P_4, \qquad 0 \leq t \leq 1,$$

where $P_i \in \mathbb{R}^2$ are control points. Control of stroke shape is achieved by varying these points and the associated tangent vectors at each anchor, e.g., $P_2 = P_1 + V_1$ and $P_3 = P_4 - V_4$ (Davydkin et al., 2023). Smooth concatenation of segments (C¹ continuity) is imposed by setting outgoing and incoming tangents to be negatives across segments.

To increase representational power and adaptability, variable-degree Bézier representations are used. Stroke degree selection is formulated as a soft discrete choice via a continuous parameter $r \in [0,1]$, enabling the synthesis of strokes ranging from linear ($n=1$) to high degree ($n \leq N$) by soft-masking basis matrices and control vectors (Das et al., 2021).

2. Decomposition of Text and Stroke Parameterization

The mapping of textual content to Bézier-curve sequences proceeds by associating each Unicode character with one or more template variants, where each variant stores anchor points and tangent vectors defining the constituent stroke segments. For each letter in the input string, a template is chosen (with page-level consistency but cross-page stochasticity), decomposed into $k$ cubic segments, and positioned according to inter-character and inter-word spacing specifications (Davydkin et al., 2023).

For cursive scripts or connected writing, segments for consecutive letters may be concatenated in a single continuous stroke; pen-lifts (disconnections) are probabilistically introduced to simulate realistic handwriting artifacts. This decomposition supports fine-grained style control at both page and segment levels.

3. Data Generation and Augmentation Pipeline

The data generation pipeline is implemented as a sequence of steps mapping text to raster or vector output. The process, outlined in pseudocode (Davydkin et al., 2023), includes:

1. Style Initialization: Random sampling of global parameters such as skew, character and word spacing, and stroke width.
2. Character Mapping: Selection of a template letter, then per-segment perturbations including:
   • Anchor and Tangent Perturbation: Jitter (Gaussian noise) on control points and rotation/length variation for tangent vectors.
   • Vertical Offset & Skew: "Dancing-line" effect (vertical noise within bounds) and affine skew matrix application.
3. Cursor Progression and Pen-Lift: Controlled advancement of the writing cursor, probabilistic stroke breaking.
4. Rendering: Rasterization at specified resolution and stroke width using libraries such as Pillow or Cairo.
5. Performance: On single-core CPUs, typical generation rates are ~100 words/sec at 300 dpi, scalable to millions of samples by parallelization.

A representative pseudocode fragment is presented in the data (Davydkin et al., 2023).

4. Controlled Augmentation and Stylistic Variability

Augmentation parameters are crucial for generating realistic and diverse handwriting. They are grouped into page-level (global) and segment-level (local) controls (Davydkin et al., 2023). Key parameters and their ranges include:

Parameter	Typical Range	Effect
σ_point	0–2 px	Stroke jitter
σ_rot	±1°–±5°	Tangent rotation noise
σ_len	0–0.10	Vector-length scaling
skew	±2°–±5°	Global slant
σ_char_std	0–2 px	Inter-letter spacing noise
σ_space_std	0–5 px	Inter-word spacing noise
y_delta_max	1–3 px	Vertical “dancing”
p_disconnect	0–0.2	Probability of pen lift

Increasing augmentation yields more "organic" results but may reduce legibility if exaggerated. A plausible implication is that overly strong augmentations degrade both perceptual realism and HTR performance.

5. Vectorization, Model Architectures, and Inverse Graphics

Generative models such as Cloud2Curve extend the Bézier paradigm to the direct synthesis and vectorization of handwriting from raster images or point-clouds (Das et al., 2021). The core ideas include:

• Inverse-graphics stroke fitting: Given a point-cloud stroke, a Bézier stroke $(\mathcal P, r)$ is optimized to minimize a composite loss (e.g., Sliced-Wasserstein distance plus regularization and MSE if available).
• Variational Autoencoders (VAE): Latent codes $\mathbf{z} \sim \mathcal{N}(\mu, \Sigma)$ encode strokes, and a decoder (LSTM) produces sequential Bézier parameters with probabilistic choices of degree and end-of-stroke.
• Deterministic Vectorization: During inference, mean estimates of the latent and sequence-decoded Bézier parameters yield crisp SVG-style output for new sketches.

The generative pipeline allows both the synthesis of entirely new handwriting styles and the vectorization (conversion) of existing raster or point-cloud samples, facilitating efficient data generation and scalable annotation tasks. Cloud2Curve demonstrates that generated samples attain higher classification accuracy (e.g., 80% on Quick, Draw! compared to 70–75% for prior SketchRNN/BézierSketch) and use fewer control points per stroke (Das et al., 2021).

6. Evaluation of Synthetic Handwriting Realism and OCR Error Patterns

The realism of synthetically generated handwriting is assessed through both qualitative and quantitative metrics (Davydkin et al., 2023):

• Visual Comparison: Bézier-generated images exhibit natural curvature and realistic stroke jitter, outperforming both font-based and neural keypoint renderings in human evaluation.
• OCR Error Statistics: When training HTR models on synthetic data, distributions of insertion, deletion, and substitution errors closely align with those from real handwritten corpora. Word Error Rate (WER) on held-out synthetic data is within 5% of WER on real HWR200 scans, both before and after POC.
• Ablation Studies: Removing augmentations such as point_noise and vector_rotation leads to an 8% reduction in HTR error but results in strokes that are visually unnaturally uniform.
• Metrics on Generative Models: Cloud2Curve achieves SWD+MSE test losses of $P_i \in \mathbb{R}^2$0 and downstream classification accuracy of 84% (compared to real data at 92%) (Das et al., 2021).

7. Applications in HTR, POC, and Beyond

Synthetic datasets generated by Bézier-based engines enable the training and evaluation of advanced sequence-to-sequence correction models for handwriting, including large architectures such as pre-trained T5 adapted for post-OCR correction tasks with substantive symbol contexts (e.g., 90-symbol input windows) (Davydkin et al., 2023). These datasets can be used for:

• POC Model Training: Accurate correction of OCR errors in handwritten text using LLMs trained with synthetic error distributions.
• Teacher Feedback and Student Assessment: Highlighting errors for educational feedback by comparing pre- and post-correction transcriptions.
• Benchmarks for HTR: Providing challenging, large-scale corpora with controlled error patterns for benchmarking modern HTR systems.

Furthermore, generative and vectorization models enable applications in style transfer, digital calligraphy, and scalable annotation tools for handwriting datasets, with open-source implementations facilitating reproducibility and further research advancements (Davydkin et al., 2023, Das et al., 2021).