An Automatic Video Text Detection, Localization and Extraction Approach

摘要

Text in video is a very compact and accurate clue for video indexing and summarization. This paper presents an algorithm regarding word group as a special symbol to detect, localize and extract video text using support vector machine (SVM) automatically. First, four sobel operators are applied to get the EM(edge map) of the video frame and the EM is segmented into Nx2N size blocks. Then character features and characters group structure features are extracted to construct a 19-dimension feature vector. We use a pre-trained SVM to partition each block into two classes: text and non-text blocks. Secondly a dilatation-shrink process is employed to adjust the text position. Finally text regions are enhanced by multiple frame information. After binarization of enhanced text region, the text region with clean background is recognized by OCR software. Experimental results show that the proposed method can detect, localize, and extract video texts with high accuracy.