Design And Implementation Of Subtitle Extraction And Recognition In Web Video

With the advancement of network service and infrastructure, the web video service has become more and more convenient, and the content of videos have been greatly enriched. Because of the variety of web video, it’s necessary to find an effective way to supervise and retrieval web videos. So that it is a better choice to understand the semantics of video, when titles and tags didn’t match the content of video always. Among plenty features of video, the subtitles strongly relate to semantics of web videos, to extract and recognize them may lead to an in-depth retrieval and supervise.The subtitles are the text content added into video manually. They can provide lots of details about the main ideas. But due to the low resolution of web video and the complicated background of text, it is not possible to use the OCR technology to recognize the character directly.To solve this problem, the author design and implement a subtitle extraction and recognition system, which can localize the subtitles in video, separate it from background, and then turn them into character code which can be understood by computer. There are three modules in the system.Subtitle extraction module has two steps:locating and tracking. Locating is used to locate the text when it first appears. The author uses the stroke width transformation to locate text region in video frame. Then a text tracking procedure is applied to video which can detect the stay and disappear of subtitles. Tracking not only reduce the cost of locating, but also increase the accuracy of subtitle extraction.The next step is the binarization and strengthening of the subtitles. To strengthen the text in image, the author integrates several images into one, so that the pixel of character can be more distinct and the background of image will be smoother. Afterwards, this paper applies Otsu to convert enhanced image into a binarized image.The last step of our system is character recognition. In this module, we have designed and extracted several features from the Chinese character, and present a cascade classifier that has three stages to recognize Chinese character. After the test, this module has been proved functional and effective.This system has been verified that it can extract and recognize subtitles in videos, and has achieved an acceptable result.