Deep learning for content-based indexing of TV programs

Abstract

[en] In recent years, deep neural networks have been successful in a lot of tasks in both industry and academia due to its scalability to mange large volumes of data and model parameters. Unfortunately, creating those large models and use their predictions can be computationally expensive to deploy on devices with limited resources. There is a TV channel called TV3 that wants to improve its recommendation engine without the mentioned impediments. In that thesis, we aim to solve part of that problem by using YOLO and Places to detect objects and scenes respectively, and build a smaller model able to learn from them and extract frame objects and scenes by itself. To do it, we have analyzed in depth Heterogeneous Classifiers (HC), that ensemble models with some different classes in a smaller model using a convex optimization approach. As HCs do not handle an scenario where classes differ completely between models, which is the TV3 case, we have implemented the smaller model following a label prediction approach by using RMSE and we have evaluated the model with ranking metrics as we have faced an unsupervised problem.