Online Handwriting Generation Using Stroke Transformation and Character-Independent Point Classification

Existing handwriting generation has seen image-based offline methods improve in quality, but background noise and unnatural stroke connections remained issues, while online methods also struggled to naturally create the ligatures that are essential to Latin and Cyrillic cursive handwriting. This method learns structure from data instead of character-specific rules, and based on the results, transforms strokes to create naturally connected cursive handwriting.

The core pipeline consists of two stages.

• Structure Segmentation: Each character is linearly resampled into 20 points, and the input consists only of coordinates and pen state. Character labels are not used; a lightweight RNN point-wise classifier using a 2-layer BiGRU (about 12K parameters) classifies each point as head/body/tail/isolated.
• Stroke Transformation: The endpoints of transformable segments are fixed, while connection distance, local smoothness, and displacement relative to the original are jointly optimized. This problem is converted into a tridiagonal linear system and solved in O(N).

Experiments were validated through English sentence generation. HWD, FID, and KID showed that the connected text was closer to the user's handwriting, and CRR/WRR showed that readability degradation after adding connections was limited to about 1%. Across the character recognizers used, the readability gap between user handwriting and generated text was also under 5%.

In a comparison of 55 image pairs by 18 users, 87% rated readability as non-inferior, and 36% rated it as better. 71% rated visual completeness as equal to or better than the original, and remaining issues mainly stemmed from individual glyph synthesis artifacts rather than stroke connections.

Efficiency was confirmed on a single thread of the Samsung Galaxy S25 CPU. The proposed method showed computational efficiency suitable for real-time editing on low-spec mobile devices, and could naturally connect or extend original handwriting at a level indistinguishable to both humans and automated tools. Ultimately, it demonstrated up to 3.21% improvement in WRR, along with the potential for character-level correction and completion of user handwriting. Future work includes extension to diverse handwriting styles and right-to-left scripts.