Modelo arquitetural baseado no NBMP para workflows de mídia com suporte a funções de IA

Abstract

The MPEG-NBMP standard (Network-Based Media Processing, ISO/IEC 23090-8) defines a model to describe and execute media processing distributed over a network,

but it does so as a lengthy, essentially textual normative document, without a reference architecture that is directly implementable. In addition, reusing processing functions especially those based on Artificial Intelligence (AI) - is difficult, since they are often developed coupled to a single flow. This work proposes an architectural model derived from NBMP for executing media workflows with support for the reuse of AI functions. The standard was translated into a class model in UML notation, later materialized as a proof of concept: a generic orchestrator, implemented in Python with Docker, that reads the NBMP descriptors and executes the described pipeline, with no logic dedicated to each case. The architecture was assessed through two use cases from distinct domains

- facial recognition in video and reading numbers on a seven-segment display -, both with core functions based on AI. The same orchestrator executed both cases without

any conditional branch per use case, with all the difference between them living in the descriptors, and integrated the AI functions as ordinary NBMP functions. It is concluded that the model derived from NBMP is executable and generic, and that the separation between orchestration and description allows AI functions to be recomposed in different pipelines without changing the code.