Video Coding for Deep Learning-Based Machine-to-Machine Communication

点击翻译按钮获取中文摘要

In this research project, the video coding for machines (VCM) task shall be researched with a strong focus on the recently emerged neural compression networks (NCNs). There, autoencoder networks are trained to reconstruct the input image from a latent space requiring as as little bitrate as possible to transmit. For the follow-up project, state-of-the-art NCNs optimized for the human shall serve as a starting point to base own optimizations and designs for the VCM task on. To that end, the project is divided into two phases. First, intra compression is researched by developing novel training losses representing the evaluation network at the decoder side in order to improve the coding gains for VCM. Thereby, it will be differentiated between generally optimizing for multiple evaluation networks and when the specific evaluation network is known and available before encoding. Second, the found methods from the first phase are supposed to be used to further optimize suitable NCN architectures designed for inter coding. However, suitable pristine labeled video data has to be acquired for an appropriate evaluation before. Besides, also test cases requiring video data such as tracking shall be considered to measure the coding efficiency.