Aceleração do particionamento de quadros intra no codificador Versatile Video Coding (VVC) utilizando redes neurais profundas

Abstract

Intra-frame prediction plays a very important role in current video coding technologies. Therefore, the current state-of-the-art standard in video coding, High Efficiency Video Coding (HEVC), the new Versatile Video Coding (VVC) standard has added new prediction modes along with new types of partitioning for intra-frame prediction. The Multi Type Tree (MTT) enables an improvement in VVC over the quadtree partitioning structure of the HEVC standard, generating more flexible block sizes. These new techniques achieve high coding efficiency, but also result in a very high computational cost, sometimes even preventing the adoption of this new standard in the industry. To mitigate this problem, this dissertation presents a fast algorithm for the decision of Coding Tree Units (CTU) partitioning during intra-frame coding, enabling a reduction in the computational cost of the process. One of the contributions of the work is to form a dataset with data in significant quantity and plurality, in order to feed a deep neural network architecture capable of classifying CTU partitioning. The experimental results showed that the proposed neural network achieved a reduction in coding time of up to 21.41%, with a loss in coding efficiency of 0.432%.