TEXT AND MOVING OBJECTS SEGMENTATION IN VIDEO FILES

Abstract

In this thesis, we focus on addressing three problems related to information extraction in the video. The first problem is fast segmentation for the “dominant” object in the video file, the second is the extraction of text caption from news bars, and the third is finding key frames related to advertisements in videos. In the first part, we introduced the use of downsampling video frames to reduce the computation time of video object segmentation while maintaining a very high segmentation accuracy. In the second part, we proposed a system to extract text caption in the news bars for different Arabic TV channels. Using optical flow and Hough transforms, the system was able to detect and classify text region into three categories: horizontal scrolling, vertical scrolling and static region. For horizontal scrolling text, the system constructed complete sentences even if the sentence spans more than one frame.In the last part, we introduce a very fast search algorithm for advertisements in the videos. The algorithm first indexes the video using some video features in a KD-tree, then a binary search is performed on the tree. All algorithms were tested, and the accuracy of the algorithms showed their suitability for practical applications.