APPLICATION OF SEMANTIC AND OBJECT SEARCH FOR OPTIMIZING MEDIA SEARCH IN OSINT INVESTIGATIONS

Abstract

The modern information space is characterized by growth in volumes of diverse media data, which has become particularly acute after the widespread adoption of large language models and increased user activity on social media. This process of growth creates fundamentally new challenges for professionals in the field of Open Source Intelligence (OSINT), who must promptly process and analyze big arrays of textual, audio, photo, and video information. At the same time, the quality of this data does not always improve proportionally to its quantity. This creates need for the development of new approaches for effective filtering and searching of relevant information under conditions of information overload. One of the tools that can help in this regard is computer vision. Computer vision allows for the automation of visual content analysis processes and the identification of key elements in media data. The application of computer vision methods combined with modern machine learning algorithms opens new possibilities for improving the efficiency of OSINT investigations, so they must be studied further. The article examines the challenges of searching and filtering information in the context of Open Source Intelligence amid increasing volumes of data. Existing methods of video frame comparison are analyzed, including pixel-level methods, feature-based comparison, perceptual hashing, object detection algorithms, and semantic embeddings. A combined approach that integrates semantic and object analysis for effective key frame detection in video materials is proposed. A method for evaluating frame importance has been developed based on five criteria: appearance of new object classes, disappearance of object classes, changes in object quantity, appearance of priority objects, and detection confidence level. The combined score is calculated as a weighted sum of semantic distance and object evaluation, providing balanced detection of key video moments. This approach creates an adaptive analysis system that ensures a deeper understanding of video content for OSINT investigations, effectively filtering out irrelevant data.