Shcherbinin K. Methods and models of video segmentation for metric search systems

The object of the research is the processes of perception and contextual interpretation of dynamic space scenes. The target of the research is to develop models and index results of segmentation in time and space to increase efficiency of the systems of metrical search and recognition of dynamic visual information of database on inquiries after the model. The methods of the research are the main statements of mathematical apparatus of recognition of images, artificial neural network, and the theory of sets, algebra and methods of mathematical statistics. The equipment to make the research is a PC, a camera and a video camera. The theoretical and practical results are the following: the methods and models to make the video segmentation have been developed; the conditions to eliminate the video segments have been found; the received notion of the space content of a video frame and picture as the whole can be used to solve a wide range of problems to search and interpret visual information. Scientific novelty consists in introducing the methods of discovery of changes in behaviour of multidimensional time series based on multidimensional modification of a procedure of exponential smoothing by an adopted control of characteristics that create prerequisites for discovering video segments in a sequential analysis of dynamic visual information. The further development has been given to the methods of metric search that take into consideration necessary and sufficient criterion on eliminating video segments without comparison with an inquiry providing considerable improvement of speed as opposed to earlier developed methods. The methods deleting semantic conflict between the low-level processing of video data and the perception target, different from earlier known, have been improved and provided ‘space content’ resulting in segmentation of separate frames of video sequences that creates the sufficient grounds to the valid search of video segments in the picture of object oriented areas. The thesis research has resulted in the complexes of video database software implemented into the systems of computer-aided diagnosis, particularly in the search in endoscopic findings video archives at the Institute of emergency surgery of Academy of medical science of Ukraine and in solving video search problems of video archives of coronarography at the Ukrainian Railway Central clinical hospital. The scientific statements presented in the thesis have been used in the courses at the Kharkiv national radio electronics university and proved with the Acts of implementation. Scientific theoretical and practical thesis results can be used by the organizations specialized on the problems of processing, research, implementation and use of smart systems of computer vision, systems of on-line environment video data as well as identification systems, in different industrial areas, in scientific and practical activity.