خوارزمية جديدة لاكتشاف النص الاصطناعي في مقاطع الفيديو الإخبارية
الكلمات المفتاحية:
Artificial text, Geometric properties, MSER, Scene text, OCR.الملخص
Artificial Text in videos plays an important role for automatically indexing video content, since much semantic information is carried with them. In this paper an effective algorithm is presented to detect artificial text region in the frames of news videos by utilizing the geometric properties of MSER regions.
The video is segmented into a sequence of frames and then the frames that contain a caption at most are selected to be the input to proposed algorithm. Firstly the Maximally Stable Extremal Region (MSER) is applied for each frame to locate a large number of individual text characters as MSERs regions. Then, candidate's text regions are filtered to remove non-text regions by using geometric properties with thresholds technique. To accurately detect text in the video frame, three types of geometric properties are utilized: centered of regions, mean intensity and bounding box. Finally, text regions are merged, highlighted and recognized using Optical Character Recognition (OCR) system directly.
Experimental results indicate that proposed algorithm provides 0.9445 recall and 0.8095 precision accuracy rate for selected video frames. The performance of the proposed algorithm is satisfactory for artificial text detection under different size, color and background.