Dictionary Learning based High Frequencynon-linear Prediction for Video Coding

高细节预测和重构是提高视频编码算法效率的关键。在这个项目中，它将研究字典学习（DL）方法如何超越线性信号处理理论，可以改善视频压缩的最新技术，以提供更好的预测和重建。在本文中，基于DL的方法相对于深度神经网络是优选的，因为它们的复杂度较低，使得它们也可以用于视频解码器中。在这种情况下，我们将研究DL /稀疏编码方法是如何受到编码伪影的影响。在此基础上，将开发用于编码噪声去除的DL方法。此外，该项目的目标是在运动补偿预测中利用基于DL的超分辨率（SR）进行子像素内插。此外，编码方案沿着使用SR的动态分辨率转换的概念将被建立和研究。在帧内预测的上下文中，基于字典学习的方法将被开发为传统方案的替代方案。该项目的最后一个目标是开发进一步降低基于DL的算法的复杂性的方法。

High detail prediction and reconstruction is a crucial task to increase efficiency of video coding algorithms. Within this project, it will be investigated how dictionary learning (DL) approaches going beyond linear signal processing theory can improve over the state of the art in video compression for providing improved prediction and reconstruction. Herein, DL based methods are preferred against deep neural networks, since their complexity is lower such that they can also be used in video decoders. In this context, we will investigate how DL / sparse coding methods are influenced by coding artifacts. On this basis, DL approaches for coding noise removal will be developed. Further, the project targets utilizing DL based super resolution (SR) for subpixel interpolation in motion compensated prediction. Moreover, a coding scheme along the concept of dynamic resolution conversion using SR will be set up and investigated. In the context of intra prediction, dictionary learning based approaches shall be developed as an alternative to conventional schemes. A last goal of the project is to develop methods further decreasing the complexity of DL based algorithms.