基於持續與雙階段學習的場景文字標註平台研究

Abstract

場景文字領域的任務旨在偵測與辨認生活周遭的文字。然而，在收集和標註場景文字的資料集時，需要大量的時間和精力，因此在標注過程中使用高效且可靠的工具，以及適當的安排標注流程是不可或缺的。我們針對繁體中文環境下的場景文字資料集之標注任務，建立了一個標注系統。在此系統中，我們嘗試使用持續學習的技術來快速調整自動提示模型，同時針對中文類別數過多，導致字元分布不平衡的問題引入長尾技術。此外，我們討論了不同的標注流程下，自動提示模型的性能。為了比較這些技術與流程的優劣，我們也提出了一個適用於標注系統的評估方式。最後，我們在系統中引入了基礎模型，以及分析不同工具對標注者標注時間的影響。透過這些技術的應用，將能實現更高效率的資料集標注。

Scene-text models aim to detect and recognize texts in our daily life environment. However, it requires lots of time and efforts to build a scene text dataset annotated by humans. Therefore, we often need to provide the detection and recognition results based on the currently available model to reduce the effort of labeling. To this end, we develop an annotation system for annotating Traditional Chinese scene text datasets. In this system, we try to use continual learning techniques to quickly adjust the automatic labeling tools and attempt to adapt existing long-tail techniques into our system to address imbalanced character distribution in Chinese characters. In addition, we discuss the performance of the automatic labeling tools under different annotation processes. To compare the strengths and weaknesses of these techniques and processes, we also propose a new evaluation protocol for the annotation system. Finally, we introduce the foundation model combined in our system as well as analyze the impact of different tools on the annotation time. By incorporating these techniques, we can achieve a more efficient annotation process.