Cross-domain Generalization of YOLO-based Models for Large-scale Cursive Chinese Character Classification
學年 114
學期 2
發表日期 2026-05-27
作品名稱 Cross-domain Generalization of YOLO-based Models for Large-scale Cursive Chinese Character Classification
作品名稱（其他語言）
著者 Ching Teng；Chii-Jen Chen
作品所屬單位
出版者
會議名稱 The International Conference on Recent Advancements in Computing in AI, IoT and Computer Engineering Technology (CICET 2026)
會議地點 New Taipei, Taiwan
摘要 This study investigates cross-domain generalization in large-scale cursive Chinese character classification under domain shift. A dataset of 985 character classes is constructed by manually cropping glyph images from cursive calligraphy copybooks. The data are divided into train/validation/test sets for in-domain evaluation, together with a cross-source wild dataset (N = 25,610) that simulates variations in color, illumination, and background conditions. We evaluate YOLO-based classification models with different capacities and systematically control color augmentation intensity (Base, M1, S1) to analyze the interaction between model capacity and robustness. Although all models achieve near-saturated in-domain performance (Top-1 = 1.000), substantial differences emerge on the wild dataset. The highest Wild Top-1 accuracy is achieved by YOLOv8s under Base augmentation (0.9972), while lightweight models exhibit stronger sensitivity to augmentation strategies. Experimental results show that the effectiveness of color augmentation is architecture-dependent. Moderate augmentation improves lightweight models, whereas excessive perturbation may degrade performance. These findings highlight the importance of explicit cross-domain evaluation and provide practical insights for designing robust cursive glyph recognition systems.
關鍵字 Cursive Chinese character classification, cross-domain generalization, YOLO-based classification, data augmentation, domain shift.
語言 en_US
收錄於
會議性質 國際
校內研討會地點 淡水校園
研討會時間 20260527~20260529
通訊作者 Chii-Jen Chen
國別 TWN
公開徵稿
出版型式
出處