講演抄録/キーワード |
講演名 |
2022-12-16 16:45
Training Kindai OCR with parallel textline images and self-attention feature distance-based loss ○Le Duc Anh・Kitamoto Asanobu(Center for Open Data in the Humanities) PRMU2022-56 |
抄録 |
(和) |
(まだ登録されていません) |
(英) |
The modern Japanese documents in late 19th and early 20th century are called Kindai documents and have great historic value for historians and experts in exploring social aspects, lifestyles, even weather in the previous era. It is time-consuming and labor-intensive work for making transcriptions for the documents. As the result, the training dataset is small and it is hard to enlarge the training dataset. In this research, we aim to enlarge small training set by parallel textline images. Parallel textline images contain a pair of original Kindai and current Japanese fonts. We propose a distance-based objective function to minimize the distance between the self-attention feature of parallel textline images. The experiments show that the proposed system improves 2.3% of CER to compare with a Transformer as a baseline Kindai OCR. Moreover, our proposed method provides a better discriminant of self-attention feature. |
キーワード |
(和) |
/ / / / / / / |
(英) |
Kindai OCR / self-attention feature distance-based loss / parallel textline images / / / / / |
文献情報 |
信学技報, vol. 122, no. 314, PRMU2022-56, pp. 127-131, 2022年12月. |
資料番号 |
PRMU2022-56 |
発行日 |
2022-12-08 (PRMU) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
PRMU2022-56 |