RESEARCH OF METHODS OF SEARCHING KEY FRAMES IN VIDEO FLOW WITH THE USE OF NEURAL NETWORKS FOR SEARCH SYSTEMS

Abstract

The paper presents a comparative analysis of current research in the field of data analysis in the format of video content and regarding it, that effective methods of analysis of these data are methods of search keyframes in the video stream. The analysis shows that the application of a method of processing visual data is determined by the structure of this data. Therefore, in order to simplify their analysis, they were divided into the following categories: consistent comparison; global comparison, based on clustering, and those that use events or objects. Especially valuable are the methods of comparing and matching matches (fragments), namely: methods of sequence search (detection of objects or certain actions on frames); methods of classifications (determining the content of personnel and their distribution to certain categories); frame decoding methods (description of the characteristics of a particular image) and methods for detecting anomalies in the video stream (search for objects, characters that are unique properties of the fragment relative to others). It shows that the most optimal of the considered methods there are methods that are based on technologies of artificial intelligence and machine learning. And also shows the difference and efficiency of deep learning methods in relation to сonventional methods. Particularly promising are the methods, the implementation of which is to model the temporal dependences of the variable range using convolutional neural networks and functions with special attention mechanisms. Methods that use an Actor-Critic model embedded in a Generative adversarial network have also demonstrated their effectiveness. It is shown that the development of these methods contributes to the rapid development of information systems with which you can successfully analyze video content and recognize its origin.