Study On Character Detection Algorithm For Video Based On DCT Domain

Extracting the semantic content of image/video automatically is of great importance in the application of image/video retrieval. Characters contained in the video are an available and important clue in content-based video retrieval and browsing. Therefore, detecting and recognizing characters automatically are the basis and key of extracting video semantic content.Currently, the detection and recognition of video images characters have done more in pixel domain, and have made some progress with application value. However, because of the volume of data, video usually storages and transmits in the form of compressed video stream. If we detect and recognize characters after fully decode video stream to a rate still images, it will result in large computation and also seriously affect the processing speed. If deal with the video stream directly after restore some of the information, then detect and recognize characters, it will reduce the total decoding time and the processing data will be also less, so it can help to improve the real-time of video retrieval.Integrating with character detection technology have been existent, this paper proposes a character detection algorithm which base on DCT-domain. It utilizes the directional features of the texture of character blocks and the property that the characters in video usually distribute in row or column. The character/non-character blocks are effectively separated by a new adaptive threshold. Then it uses morphological operation to smooth and filter the binary image and verify the candidate text regions. Finally, the text regions are accurately obtained by horizontal and vertical projection. Experimental results demonstrate that the proposed algorithm can detect characters accurately even in video with complex background, and it is of good robustness and high practical value.Point at the text region which has been located, a new image binaryzation algorithm which merges image histogram statistics and edge feature is proposed simultaneously. First it uses the image gradient operator to detect edge points. Then it chooses the highest and the lowest intensity points within every edge point's 8-neighborhood as the high and the low intensity points, respectively take the mean of the high intensity points as the high threshold, and take the mean of the low intensity points as the low threshold, then counts histogram about all pixels of the image whose intensity are between the high threshold and the low threshold. Finally, the histogram is used as a parameter of otsu algorithm to binarize the character image. The binarization algorithm has lower complexity and integrates with global histogram information and edge information, so it achieves nice segmentation results.