Abstract:
Information extraction from multimedia content is a challenging task.
In this thesis, we present an architecture of multimedia contents classification system that provides different phases to extract semantic information from broadcasted
streams, starting with the segmentation process, news topics extraction, and advertisement detection and classification. Next, we give an extension to our framework
and describes an audio-based hybrid model for content classification combining different deep neural networks with auto-encoder applied to advertisement detection in TV
broadcast. Our models achieve high levels of precision. The last contribution consists
of a distributed architecture based on the Kafka and Spark frameworks which offer
parallel processing of TV streams, we demonstrate through this work the scalability
and robustness of this architecture.
