Research On Ultra-low Quality Video Caption Text Recognition

Nowadays, there is strong demand for analyzing information in video data because of abundant information in video. Some videos have very low quality due to various reasons. We define these videos are “ultra-low quality videos”. Traditional video text recognition methods often fails to process ultra-low quality video appropriately. So it is necessary to research the method of ultra-low quality video text recognition. In this article, the recognized text belongs to caption text, with fixed position and uncomplicated background.The ultimate goal of this articleis to build a video text recognition system with high performance and reliability. Firstly we purposed the template-based character recognition method, using image gray value as feature to calculate similarity of two images. And take the highest similarity class label as the recognition result. Take into account the practical requirements, the method of automatic video text area recognition was purposed. After using various template library to recognize the same frame, it generates average credibility of recognition result for each template library. Finally the text area with highest credibility was selected as recognition result of current frame.To reduce the user huge workload of establishing a template library, the automatic template library generation method was purposed. And this method based on semi-supervised self-training algorithm. Firstly we generate the standard templates with known label as origin template library. Secondly the statistical similarity distribution between each standard template and video character images was calculated. Finally the top k characters in the similarity distribution was selected, added to current template library. After above steps the original template library was generated.When the font of recognized video text is unknown, there will be some errors in auto generated template library inevitably. And it affects the recognition accuracy. To solve this problem, the method of recognition result post process algorithm was proposed. It includes the auto correction algorithm based on autoregressive model and the artificial error correction algorithm with feedback mechanism. The possibility of nature value mutation is low, based on that assumption, the autoregressive algorithm can detect the mutation points and correct them automatically. Artificial error correction algorithm let user label the right recognition result directly, and add the corrected image to template library, at the same time we correct the error in template library. After the manual correction, video was recognized iteratively using new template library. And the accuracy rate increased in each iteration.The experiment indicated that using our methods to recognize ultra-low quality video text, after 3 correction iteration finished, the average recognition accuracy rate can reach more than 98%. Finally we designed and developed a video character recognition system. And the system has friendly user interface and high reliability.