- 學位論文
基於RGB影像的空中手寫繁體中文字辨識應用研究
A Study and Application of In-air Handwriting Traditional Chinese Character Recognition Using RGB Image
摘要
隨著科技進步，科技裝置貼近人類的生活，除一般電腦以外，手機、平板及智慧手錶等電子設備如雨後春筍進入每個人的生活中。更隨著時代的推進，電子設備從原先一開始一整個房間大小的電腦，變成了可戴在手上隨身的手錶，便利性與隨身性日漸重要，也漸漸突顯了一個議題：輸入電子設備的方法。 在疫情爆發後，民眾對於個人及環境上的衛生議題逐漸重視，會有多人接觸到的共用鍵盤或是觸控板等等提高了傳染的風險，而不會接觸到輸入設備的語音輸入及空中手寫辨識的重要性也日漸提升。相較於語音辨識，空中手寫擁有更佳的隱私保密性。 而在空中手寫時，因為無法有效的分辨使用者的「提筆」與「下筆」的動作，書寫完畢時字跡會全部相黏在一起，這些在原本平面書寫時不應該出現的字跡，稱之為「虛筆」。而繁體中文正方形的字體有著較於繁雜的筆劃數量與筆劃順序的特性，在虛筆發生時會相較英文及簡體中文更加的影響辨識的正確性，使得繁體中文手寫的辨識模型建立困難。 本次研究為改善在空中手寫時產生虛筆的問題，以ROI(region of interest)方式加強「提筆」與「下筆」分辨能力，以消除書寫時虛筆產生，使輸入辨識模型時能貼近於平面上手寫的狀況，讓空中手寫辨識繁體中文字的模型可以使用一般手寫字體建立模型，而不需再另外收集空中手寫的數據，減少數據收集時間及建模參數且可維持該模型在空中手寫輸入上的辨識能力。本次更加入手部姿勢辨識指令，使用者除了在書寫時可命令開始書寫及結束書寫，更加入了可以在寫錯時，修正寫錯筆劃的功能，大幅減少以往空中手寫時需一筆完成的壓力，提高使用者的書寫體驗 最後我們設計十題中文手寫題目並請三名試驗者進行空中手寫實驗，比較輸入時有虛筆及無虛筆時對於相同手寫體模型正確率影響，我們得到無虛筆90%的正確率，高出有虛筆37%。在穩定性測試上，有虛筆正確率發生20%的浮動，而無虛筆仍穩定在90%以上。
並列摘要
With the advancement of technology, electronic devices such as smartphones, tablets, and smartwatches have become ubiquitous in people's lives, in addition to traditional computers. As technology has progressed, electronic devices have evolved from the large room-sized computers of the past to the portable devices we carry with us today, highlighting the importance of convenience and portability, as well as the issue of input methods for these devices. Following the outbreak of the pandemic, people have become increasingly aware of personal and environmental hygiene issues. The use of shared keyboards or touchpads that are touched by multiple individuals has increased the risk of transmission. Therefore, the importance of voice input and air handwriting recognition, which do not require direct contact, has become more prominent. Compared to voice recognition, air handwriting offers better privacy and confidentiality. However, in air handwriting, since it is difficult to distinguish the user's "upstroke" and "downstroke" movements, the written characters tend to stick together, resulting in what is known as "virtual strokes". Traditional Chinese characters have more complex stroke sequences than English or simplified Chinese, and virtual strokes have a greater impact on their recognition accuracy, making it difficult to build air handwriting recognition models for traditional Chinese characters. To address this problem, this study uses the region of interest (ROI) method to enhance the ability to distinguish "upstrokes" and "downstrokes" in air handwriting, eliminating the virtual strokes that occur during writing. This enables the air handwriting recognition model to closely mimic the conditions of handwriting on a flat surface, allowing the use of regular handwriting fonts to build models for recognizing traditional Chinese characters without the need to collect additional air handwriting data. This reduces data collection time and modeling parameters while maintaining the model's recognition capabilities for air handwriting input. In addition, this study incorporates hand posture recognition commands, allowing users to command the start and end of writing and to correct mistakes made while writing, significantly reducing the pressure of completing a stroke in a single motion and improving the user's writing experience. Finally, we designed ten handwritten Chinese character tasks and asked three test subjects to perform air handwriting experiments. We compared the accuracy of the same handwriting model with and without virtual strokes. We obtained an accuracy of 90% without virtual strokes, which was 37% higher than with virtual strokes. In stability testing, the accuracy with virtual strokes fluctuated by 20%, while the accuracy without virtual strokes remained stable at 90% or higher.