網路圍棋比賽影片影像自動辨識系統之開發

Development of an automatic recognition system for online Go (Weiqi) videos

緒論：本研究旨在通過開發一套可自動辨識與分析網路圍棋比賽影像的系統，達到自動提取和辨識棋盤上的訊息，做為未來分析圍棋比賽內容的基礎分析工具，以提升我國圍棋選手的能力與國際圍棋賽事的成績。方法：研究者選定YouTube上的公開比賽影片作為數據來源，並設計了四個主要的功能模組：影像轉換模組、棋盤切割模組、棋盤取子模組和棋盤串聯模組。首先，影像轉換模組使用網頁爬蟲技術和YouTube API獲取影像，並利用OpenCV將其轉換為圖像幀。接著，棋盤切割模組通過DETR模型進行棋盤檢測和信息提取，並採用影像增強技術提高模型的泛化能力。在棋盤取子模組中，研究者將棋盤劃分為361個等分，並使用ViT模型進行棋子分類和檢測，從而減少標記數據需求。最後，棋盤串聯模組整合不同時間序列的棋盤信息，開發了提子和時序轉換算法，實現了棋局的連貫記錄。結果：本研究蒐集了135支來自YouTube的圍棋比賽影片，並依據影片的完成度進行篩選，篩選後保留了86支具有完整比賽過程的有效影片。影片總長度為1,281,642.77秒，共產生1,332,791張圖片。經由本研究的研究方法進行辨識判定後，本研究從每支影片中隨機抽取約6張圖片，使用人工審核檢視研究結果的正確性，並採用嚴格的正確率公式來判定影像辨識的正確性，依據驗證的結果，影像判斷的正確率為82.56%。結論：綜上所述，本研究成功實現了自動提取和辨識圍棋比賽視頻中的棋盤信息，為圍棋數據分析提供了堅實基礎。然而，系統的整體性能仍有提升空間，未來的改進將進一步增強其實用性和泛化能力。

Introduction: This study sought to automatically extract and identify information on the Go board by analyzing online Go tournament videos. Methods: The researchers selected publicly available tournament videos on YouTube as the data source, and designed four main functional modules: a video conversion module, a board segmentation module, a stone detection module, and a board sequence reconstruction module. The video conversion module used web scraping techniques and the YouTube API to acquire the videos, which were then converted into image frames using OpenCV. The board segmentation module employed the DETR model for board detection and information extraction, utilizing image enhancement techniques to improve the model’s generalization ability. In the stone detection module, the researchers divided the board into 361 equal parts and used the ViT model for stone classification and detection, thus reducing the need for labeled data. Finally, the board sequence reconstruction module integrated board information across different time sequences, using algorithms for stone removal and sequence transformation to achieve coherent game record reconstructions. Results: The study collected 135 Go match videos from YouTube, which were then filtered based on the completeness of the matches. After filtering, 86 videos with complete matches were retained. The total length of these videos was 1,281,642.77 seconds, and they generated a total of 1,332,791 images. Using the research methods developed in this study, approximately six images from each video were randomly sampled for manual review in order to verify the accuracy of the results. A strict accuracy formula was used to determine the correctness of the image recognition, and the validation results showed an image recognition accuracy of 82.56%. Conclusion: This study successfully achieved the automatic extraction and identification of Go board information from tournament videos, providing a solid foundation for Go data analysis. However, there was still some room for improvement in the overall performance of the system. Future enhancements will further increase its practicality and generalization capabilities.