Vers le déploiement du codage vidéo neuronal appris de bout en bout
Marwa Tarchouli · theses.fr (ABES) · 2024
In recent years, there has been an exponential surge in video consumption, increasing the demand for superior video quality and high resolutions. This has motivated efforts to enhance video coding algorithms. Hence, the latest traditional codec VVC was developed in 2020, boosting the rate-distortion performance by 50 % compared to its predecessor HEVC . On the other hand, the popularity of Deep learning algorithms is growing rapidly, earning acclaim for their outstanding performance in numerous fields including image processing. This recognition captures the interest of both the video compression and the artificial intelligence communities. These factors contribute to the emergence of learned compression. This thesis explores end-to-end learned video compression systems, focusing on two main methodologies: VAE-based codecs and INR-based codecs. While VAE-based codecs surpassed the performance of the latest codec VVC, their heightened complexity presents challenges for practical deployment in the industry. On the contrary, INR-based codecs, are designed to maintain low complexity and incorporate content-adaptive networks. Nonetheless, they did not reach yet a sufficient coding efficiency level to rival handcrafted video codecs. In this thesis, a first contribution is proposed for VAE-based codecs to mitigate the increased computational complexity issue. Furthermore, a second one is introduced to enhance the coding efficiency of INR-based codecs.