Caption Recognition Researches Based On DM8168

In the age of massive data, people are trying to locate the video content they are interested in, which is however, becoming increasingly hard with the explosive growth of video data. Moreover, the surveillance of legal use and distribution of videos also faces great challenges in China. In many cases, the captions of the videos can be useful resources for the understanding of both the contents and meanings of online videos due to the semantic relevance between them. Therefore, the automatic recognition of Chinese captions can largely contribute to the endeavor of content regulation and retrieval.Research in this paper is based on embedded platform with the state-of-the-art DSP processor DM8168. Compared with PC, Embedded platforms is more convenient, less expensive, but also more difficult to develop. Research work of this paper is divided into two pieces:one is to be familiar with the hardware platform and to build the video data link; another is the development, porting and optimization of caption recognition algorithmsCaption recognition algorithm mainly includes the caption extraction and characters recognition. After analyzing the characteristics of the video captions, this paper proposed the caption extraction algorithm based on the contrast of images, using the convolution project method to locate the caption position. About the character recognition algorithm, he post processing of the binary phrases is also used to improve the accuracy of character recognition.Experiments show that, the caption recognition algorithm proposed in the paper can run real-timely and stably on the embedded platform, and can process various caption styles, which demonstrates the robustness and adaptability of the algorithms. As to the performance, the algorithm achieves more than90%accuracy in caption detection, and at least80%accuracy in character recognition.