Liangjiang enterprise triumphs in handwriting recognition at ICDAR 2026

english.liangjiang.gov.cn|Updated: 2026-09-11

文字识别_副本.jpg

Handwriting recognition algorithm expert Qin Xunhui from Chongqing Isigning Technology Co receives awards at ICDAR 2026. [Photo provided to english.liangjiang.gov.cn]

Chongqing Isigning Technology Co, based in Liangjiang New Area, clinched three world championship titles in the handwriting recognition category at the 20th International Conference on Document Analysis and Recognition (ICDAR 2026), held from Aug 30 to Sept 4 in Vienna, Austria.

Isigning demonstrated exceptional prowess by securing top positions in all three handwriting recognition competitions: AnyScript for long-term handwriting author identification, FalsID for historical manuscript falsification and imitation detection, and CircleID for handwritten identification.

In the AnyScript competition, Isigning's advanced modeling of micro handwriting features and macro writing habits enabled them to decode hidden "handwriting DNA," securing their victory.

In the FalsID competition, Isigning's AI algorithms surpassed human capabilities in distinguishing authentic from fraudulent manuscripts. Their proprietary deep elliptical encoding model outshone global research institutions and tech giants, achieving the championship with an impressive 11-point lead over the second-place team.

Isigning's paper, AOSSig4000: A New Chinese Handwritten Signature Dataset with Diverse Background Noise and Pixel-Level Annotations, was accepted at ICDAR 2026. This paper introduces the world’s first real-scene Chinese handwritten signature dataset with pixel-level annotations, "AOSSig4000," accompanied by the AOSDoc document template set.

The dataset comprises 4,000 documents and 8,523 handwritten signatures, systematically incorporating real noise and categorized into five difficulty levels (L1 to L5). Researchers meticulously annotated each signature over a total of 2,000 hours.

This research highlights the superior benefits of real data over synthetic data for model training, showing that using signature segmentation as a preprocessing step significantly reduces error rates in complex interference scenarios.