Towards novel inter-prediction methods for image and video compression
Jean Bégaint · ORI-OAI warehouse (University of Rennes) · 2018
Due to the large availability of video cameras and new social media practices, as well as the emergence of cloud services, images and videosconstitute today a significant amount of the total data that is transmitted over the internet. Video streaming applications account for more than 70% of the world internet bandwidth. Whereas billions of images are already stored in the cloud and millions are uploaded every day. The ever growing streaming and storage requirements of these media require the constant improvements of image and video coding tools. This thesis aims at exploring novel approaches for improving current inter-prediction methods. Such methods leverage redundancies between similar frames, and were originally developed in the context of video compression. In a first approach, novel global and local inter-prediction tools are associated to improve the efficiency of image sets compression schemes based on video codecs. By leveraging a global geometric and photometric compensation with a locally linear prediction, significant improvements can be obtained. A second approach is then proposed which introduces a region-based inter-prediction scheme. The proposed method is able to improve the coding performances compared to existing solutions by estimating and compensating geometric and photometric distortions on a semi-local level. This approach is then adapted and validated in the context of video compression. Bit-rate improvements are obtained, especially for sequences displaying complex real-world motions such as zooms and rotations. The last part of the thesis focuses on deep learning approaches for inter-prediction. Deep neural networks have shown striking results for a large number of computer vision tasks over the last years. Deep learning based methods proposed for frame interpolation applications are studied here in the context of video compression. Coding performance improvements over traditional motion estimation and compensation methods highlight the potential of these deep architectures.