Machine Learning-based Adaptive Reconstruction of Video Stream Fragments Taking into Account Scene Dynamics. Proceedings of the Institute for System Programming of the RAS

Труды Института системного программирования РАН. 2026. Т. 38. № 1. С. 255–274.

Думкин Н. А., Alexandrov D., Прозорский М. А.

A theoretically sound approach to adaptive client-side video fragment restoration is proposed using
machine learning and scene analysis methods. The method includes a formal problem statement, a finite-state
machine model for decision making, a restoration cost function, and a new stage in video preparation: scene
dynamics assessment followed by recording a feature in an HLS playlist. This feature improves the accuracy
of selecting video fragment restoration methods. The algorithm architecture is described, rules for selecting a
restoration strategy are proposed, and the use of additional features in the formalized model is theoretically
justified. The work is primarily theoretical and methodological in nature: a restoration strategy is formalized
and a playlist extension is proposed. A limited simulation evaluation is also provided, the purpose of which is
to verify the internal consistency of the model rules and sensitivity to threshold parameters, rather than to
demonstrate superiority over existing methods.

Keywords: Adaptive Video Streaming, HLS, video content restoration, video frame interpolation, video super-resolution, client-side restoration algorithms, адаптивный видеостриминг, восстановление видеоконтента, интерполяция видеокадров, суперразрешение видео, клиентские алгоритмы восстановления, стандарт HLS

Publication based on the results of: Methods and algorithms for the use of intelligent agents in the development and refactoring of distributed software systems (2026)