Mikhnova O. Model and method of key frame extraction for video summarization

The thesis is devoted to the development of such a field of science as video processing, namely key frame extraction from different video genres with the help of spatio-temporal segmentation techniques. From the point of temporal video segmentation, the benefits and drawbacks of current shot boundary detection techniques are observed. From the point of spatial segmentation, field of view partition of each frame is performed via Voronoi diagrams by taking into account color, texture and shape features of the tessellations. Method for salient point detection is perfected, under which Voronoi diagrams are built. Possibilities of higher order Voronoi diagram construction for video content presentation and their peculiarities are analyzed. Specialized metrics are developed, that provide means for comparison of Voronoi diagrams and detection of considerable changes in the whole video content. Key frame extraction method is developed based on the proposed mathematical model designed for presentation and analysis of video content. It enables not only to extract frames with significant changes in content, but also to eliminate frames with similar content. The carried out experiments confirm competitiveness of the model and method which are successfully integrated into an industrial enterprise activity.