視訊字幕之偵測與辨識

Abstract

本篇論文的主要目的在於研究如何正確的偵測、擷取和辨識中文視訊字幕，以作為視訊內容分析的基礎。本論文利用字幕邊緣呈現高對比差與文字邊緣閉合兩個特性，偵測出文字區域的位置，再決定文字的顏色，擷取出文字區塊。針對中文字常由多區塊組成，同一字之區塊必須結合以利後續光學辨識系統分析的問題，本篇論文提出一個計算區塊群屬的距離公式，將區塊依其大小與排列關係分群，使同文句之區塊分為同群，再計算同群區塊的長寬，作為區塊之間結合的依據，並透過此歷程，排除非文字區塊之雜訊，強化系統在視訊複雜背景下的穩定度。最後針對低解析度造成視訊字幕品質不穩定及字體多變的問題，我們提出一個二階段辨識模型，解決視訊文字辨識所面臨的困難。

This thesis focuses on caption detection, extraction and recognition in videos. The proposed method uses the high contrast edges and the closed-form boundaries of characters to detect and located caption regions with high precision. Connect ed component analysis is used subsequently to segment out character components. A distance measure between components is defined to guide the merging of the components from the same character and to filter out non-text noise. Finally, a two -stage classifier is proposed for Chinese video OCR with the ability to tolerate poor image quality and the presence of multiple fonts.