Research On Video OCR

Texts in video frames often carry the most important information, such as place, time, name or topics, etc. As a senior semantic information, they may do great help to video indexing and video content understanding. This paper studies the video text extraction and recognition. The main parts are the algorithm and key technology of text localization, text tracking and text segmentation.In the step of text localization, this paper proposes a new text localization algorithm for video frames in a localization-to-verification framework. In text localization module, taking full advantage of character stroke attribute, the algorithm introduces the stroke operator which has strong response to text regions; subsequently, performs strokes extraction, stroke density filtration and region decomposition to obtain candidate text boxes. In verification module, the algorithm extracts edge oriental histogram features, which have strong discriminabilities for text and non-text, then Adaboost classifier is used to verify candidate text boxes.In the video text tracking, taking into account texts in the video frame are almost rigid in static or uniform linear motion, this paper proposes a tracking algorithm based on block matching and motion estimation. In order to improve tracking speed, adaptive three-levels tracker is designed. In response to rigid motion texts, liner prediction, text region comparison and edge bitmap comparison are combined, which enables the real-time tracking and achieve accurate alignment among the pixels.Existing video text segmentation algorithms are generally based on a single attribute of video texts. This paper presents a segmentation algorithm which comprehensively utilizes various attributes of video texts, such as temporal and spatial, strokes, color, geometric. This algorithm uses multi-frame integration to enhance the images; performs the stroke operator to extract the character strokes based on the feature of the stroke width; then analyses the color of the character strokes according to the stroke image and extracts the color layer of the character; finally removes the background with the same color and noises in the color layer of the character by the improved connected components analysis.Particularly, based on the algorithm research, we implement a video text extraction and recognition demo system, which can automatically generate text-frame indexes for video files.