Codec Agnostic Strategies for Enhanced Learned Media Compression

In recent years learning-based codecs have gained a lot of momentum in the field of multimedia compression. The reason for their success is the ability to learn very informative yet entropy efficient representations of the data. This advantage is what allows them to outperform traditional codecs in most coding tasks. However, most of these codecs are still in early development and are lacking some of the features that are common in traditional codecs. Some examples are high and low complexity operation modes, quality scalability, semantic information exploitation, and, although mostly relevant when talking about learned solutions, the ability to choose if the sample should be decoded with high fidelity or high perceptual quality. In this thesis, some codec-agnostic algorithms that address some of the aforementioned problems are presented. The algorithms should be as independent as possible from the codec design and they should not impair the ability of the codec to be used in the standard way. This should allow them to be easily implemented into any new codec without requiring ad-hoc designs thus allowing to deploy efficient and flexible codecs in a short time. The algorithms proposed in this thesis follow the aforementioned principles since they increase the flexibility of the codecs they were implemented on without considerably affecting their compression efficiency and encoding/decoding time. This constitutes a step forward in the usability of the new generation of codecs in real-case scenarios.