Redução do tempo de codificação de mapas de profundidade do 3D-HEVC usando árvores de decisão estáticas construídas através de Data Mining

Abstract

This dissertation presents a solution for time reduction in 3D-High Efficiency Video Coding (3D-HEVC) depth maps coding. With the insertion of depth maps in 3D-HEVC was possible significantly reduce the size of the encoded and transmitted video. However, the depth maps should be efficiently encoded for generating synthesized views with good visual quality. In addition to the tools used to encode texture data, the 3D-HEVC adds new tools developed for encoding depth maps evaluating many block sizes and encoding modes generating a high computational cost. This work proposes an approach, which uses data mining technique for training six decision trees with information extracted from reference software 3D-HEVC Test Model (3D-HTM) 16.0. Each decision tree is responsible to decide if the evaluated Coding Unit (CU) should be divided into smaller sizes. The decision trees were constructed for Coding Units (CUs) sizes 64×64, 32×32 and 16×16 and three decision trees are specialized for I-frames and three specialized for P- and B-frames. When evaluating the solution with 3D-HTM 16.0 was possible to save 52% of total execution time with a negligible impact in Bjontegaard Delta-rate (BD-rate) of 0.18%, when compared to 3D-HTM without modifications. Besides, the results demonstrated that the solution outperformed the related works.