The Research And Implementation Of Error Correction Of Video Caption Recognition Results

Video captions help people understand the content of the video, so they play a crucial role in video files. How to correctly obtain the video captions becomes a hot research topic in multimedia technology. However, by utilizing existing image processing techniques and character recognition techniques, the caption always has a lot of errors. How to correct these errors is also a big problem needing a solution.In this thesis, we studied the automatic correction technology of OCR results of video captions, and analyzed the properties of them. Firstly, we proposed a TF/IDF-based corpus design approach, and established a keyword corpus related to video contents. Then, we proposed a three-level correction system framework. First, using a redundant information-based approach, we propose a character-level single character correction method; and then involving full separation into word-separation algorithms, we combine least risk Bayesian decision to weighting processing of words, so as to complete the word-level process. Last, based on a graph theory, we use Dijkstra algorithm to finish the sentence-level correction. Using the methods mentioned above, this thesis designed and implemented a three-level correction system for video caption recognition results. In the experiment module, we use the video data obtained from CCTV News programs as the experimental data to test the performance of the proposed system. The experiments showed that, the proposed three-level correction system could generate efficient correction result for video caption recognition results.