Subtitle Text Detection Method based on Faster RCNN

• 1、Department of Automation, Beijing University of Posts and Telecommunication, Beijing 100876

Abstract：Subtitles in videos often have strong semantic information, which can effectively help people understand and analyze video content. The purpose of this paper is to use deep learning technology to detect subtitle text in video image quickly and accurately, so as to achieve efficient and fast acquisition of video data content information, and to assist relevant practitioners in the retrieval and classification of massive video data. First, the corresponding video subtitle dataset is established. The character samples of the data set include 6763 common Chinese characters, 26 English characters, etc. The diversity, balance and generalization of the samples are fully considered. Second, aiming at the specific scene of video subtitle text detection, this paper chooses Faster RCNN detection framework, introduces transcendental loss function and other methods to improve accuracy and recall. Finally, the whole system\'s serial connection and construction are completed, from video reading to caption frame interception, then to caption text line positioning, text content detection, the end-to-end process from video input to text string output is achieved. A video subtitle detection system with both speed and accuracy is designed and implemented, which can realize real-time location and recognition of video subtitle text. The accuracy and recall rate of text detection have reached 99.5%, the top-1 accuracy rate of text recognition is 97.5%, and the overall detection speed is 45 fps.

Keywords： artificial intelligence deep learning subtitle detection text localization