Text Extraction In Video

Text in digital video can provide important supplemental information for retrieval and indexing. There are cases where text in a clip contains information that is not found anywhere else such as movie credits, and other cases where text is an important concise supplement, such as sports scores or stock prices. Many high-level applications such as video abstract become possible if text in digital video can be extracted and recognized robustly.This dissertation presents our work on several aspects of text extracting in digital video, including text localization, tracking, enhancement and segmentation. Compared with typical document images text in video presents challenges because of low resolution, complex background, lighting variation, and unrestricted pose, shape and color.A method to automatically localize texts in the compressed domain and spatial domain is presented. The text regions are detected directed in DCT domain using the texture energy of each DCT block. A horizontal projection profiled of differential image of text region is employed in text line extraction.The tracking algorithm makes use of template matching with M-estimator. The matching template is acquired by segmenting the text region using logical level technique. The location of search window is estimated by using the motion vectors in the MPEG-2 bitstream. Multi-resolution method based on the winner-update strategy is adopted to speed up the template matching.An enhancement algorithm by multi-frame integration is used to increase the contrast between text and background. We decide to adopt multi-frame averaging method or multi-frame minimizing/maximizing method to enhance the text region by the analyzing the intensity distributing of each pixel over time.A text segmentation algorithm based on color stroke model is proposed. The color stroke model depicts the local topographical feature of characters in color space. The algorithm combines the binarization of text region and connected components analysis.