利用樣本重組的半監督學習之場景文字偵測

Abstract

隨著科技的發展與進步，生活周遭所帶來的資訊越來越重要。在任何一個場景中，周遭的文字訊息都跟周圍的環境有著極高的相關性。若我們能自動偵測場景中的文字，以利後續的資料收集與分析，勢必能為生活帶來更多的方便性。場景文字偵測這項研究中，相關研究大多以英文為主。雖有少數工作研究簡體中文，但繁體中文幾乎沒有。而場景文字偵測為一般物件偵測中的一個特定應用，所以我們提出基於物件偵測的方法，應用在場景文字偵測上。而物件偵測方法大多使用監督式學習，其依賴大量的訓練樣本，但在真實世界中，標註樣本取得不易，所以聯合運用非標註樣本的半監督式學習方法，較符合真實世界的需求。本研究打造一個半監督式繁體中文場景文字偵測模型。透過交換具標註樣本的背景與文字來合成新樣本，並配合拼貼的資料增強方法，豐富訓練樣本的多樣性，實驗證明本論文提出的樣本重組能更有效地運用標註與未標註樣本。關鍵字:深度學習、物件偵測、場景文字偵測、半監督式學習。

With the development and advancement of technology, the information in our surrounding environment becomes increasingly accessible. Texts in a scene reveal a lot of information about the environment. Automatic scene text detection is essential for subsequent text recognition, understanding and analysis. However, most existing studies focus on English texts, while the annotated datasets of traditional Chinese and other languages are scarce. Since text is a specific instance of general objects, we develop a Traditional Chinese text detector based on object detection. As object detection methods are typically supervised that relies on a large number of training data with ground truth labels, we apply semi-supervised learning in this work that explores unlabeled samples, which is more practical for real-world applications. In particular, we expand the limited labeled data by data synthesis and mosaic. Experiments demonstrate the effectiveness of the proposed method.Keywords: Deep learning; object detection; scene text detection; semi-supervised learning;