Deep Learning in Video Coding

Vorschlag für eine Bachelorarbeit / Masterarbeit / Forschungspraktikum


Deep Learning in Video Coding


In the past years deep neural networks have been shown to be very efficient for image and video coding. In video coding, the signal is to be transmitted using the smallest amount of bits possible. Two major concepts from classical compression—prediction and transform coding—can be very well generalized and improved using neural networks since they are capable of further optimizing the previously linear functions used for this task.

The performance of such networks largely depends on the network structure and the training set. In both areas there is room for improvement and further research. In general, there are mainly two approaches: So-called end-to-end trained coders consist of only neural network which transform and predict the signal down to a sparse and easily transmittable representation from which the image is then reconstructed by another neural network. On the other hand, many approaches build upon current state-of-the-art classical video coders by replacing individual components with network-based systems. In particular, this often includes prediction steps and in-loop-filters.

Possible topics for a thesis include:

  • Optimizing the network structure for image coders
  • Scalable end-to-end image coding
  • Variable rate end-to-end image coding
  • Neural network-based intra prediction in classical image and video coders


Prof. Dr.-Ing. André Kaup


Fabian Brand,


Python, DeepLearning desirable

Some topics require C++


Ab sofort