CalliGUI: Interactive Labeling of Calligraphic Character Images

Abstract

Abstract-Calligraphic data entry is accelerated by generating, with a feature-based character classifier, an ordered list of reference candidate labels for each character image. The improvement of labeling throughput depends on the top-N accuracy of the classifier, which in turn is a function of the available already-labeled patterns. Experiments on a database of 13,351 ancient calligraphic characters indicate that clicking on reference labels is more than twice as fast as Pinyin keyboard entry.

The CADAL calligraphic database

“…11. It is superficially similar to that reported in [12] but contains changes in the links to the database, improved logging facilities, more consistent tracking of the characters on the page, better layout, and improved ability to accommodate the user's browser display. On the left is an image of a page in running style from which individual characters were segmented.…”

“…Experimental results on a PC platform, three subjects, and three works, reported in [12], show that label entry with CalliGUI is more than twice as fast as Pinyin keyboard entry. The assistance provided by the classifier consisted of the top 25 label candidates generated after training on the database but excluding the current work.…”

Stroke-Based Character Segmentation of Low-Quality Images on Ancient Chinese Tablet

“…However, to the best of our knowledge, few published reports have focused on automatic character localization and segmentation of low-quality Chinese tablet images as summarized in [29]. Most of the aforementioned methods are not suitable for accomplishing tablet segmentation when tablet images have (1) large cracks and significant abrasions, (2) when variable widths of Chinese character strokes are present, and (3) when large numbers of intersecting and overlapping strokes are present.…”

Optimal policy for labeling training samples

“…3, 4, 5 CAVIAR systems 6,7,8,9 , active learning 10 and interactive polygons 11 are more recent manifestations. "Unsupervised" classification methods like co-training 12,13 also make use of classifier-assigned labels for retraining, but cannot guarantee the same accuracy as human correction.…”