A novel video text extraction approach based on Log-Gabor filters
Xiaodong Huang · 2011
Video text brings important semantic clues about video content. Text extraction is a crucial stage of analyzing the video text. Most of papers perform video text extraction using stroke, intensity features, which are sensitive to the video background. Text character extraction is difficult due to the complex background of video frames. The structure of text characters exists in its phase information, which is insensitive to the background. The phase map is first generated to retrieve the phase angle image using Log-Gabor filter. Then, we propose a novel text extraction algorithm based on phase map. First, for the text row in single frame, we retrieve the phase map. Second, we perform k-means clustering in the phase map and select one clustering result as the text character image. Third, in 30 consecutive frames which contain same text character, we select the four text character images and combine them into one binary character image. Finally we use the dam point labeling and inward filling [1] to remove some noise and get the binary character image. Experimental results show that this approach is robust and can be effectively applied to text extraction in video.