Bagged Tree Based Frame-Wise Beforehand Prediction Approach for HEVC Intra-Coding Unit Partitioning

Abstract

High Efficiency Video Coding (HEVC) has achieved about 50% bit-rates saving compared with its predecessor H.264 standard, while the encoding complexity increases dramatically. Due to the introduction of more flexible partition structures and more optional prediction directions, HEVC takes a brute force approach to find the optimal partitioning result which is much more time consuming. Therefore, this paper proposes a bagged trees based fast approach (BTFA) and focuses on the coding unit (CU) size decision for HEVC intra-coding. First, several key features of a target CU are extracted for three-output classifiers. Then, to avoid feature extraction and prediction time over head, our approach is designed frame-wisely, and the procedure is applied parallel with the encoding process. Using the adaptive threshold determination algorithm, our approach achieves 42.04% time saving with negligible 0.92% Bit-Distortion (BD)-rate loss. Furthermore, in order to calculate the optimal thresholds to balance BD-rate loss and complexity reduction, the neural network based mathematical fitting is added to BTFA, which is called the advanced bagged trees based fast approach (ABTFA). Finally, experimental results show that ABTFA achieves 47.87% time saving with only 0.96% BD-rate loss, which outperforms other state-of-the-art approaches.

Transformer-Based Data-Driven Video Coding Acceleration for Industrial Applications

“…To further analyze the coding performance of the proposed algorithm, we compare it with three recent studies. Respectively, they are convolution neural network-based fast algorithm (CNNFA) proposed by Liu et al [18], fast intra-CU splitting algorithm (FICUSA) proposed by Zhong et al [30], and bagged tree-based fast algorithm (BTFA) proposed by Li et al [31]. CNNFA, FICUSA, and BTFA are all effective schemes for intra-CU size decision of HEVC, and their performance on standard test sequences is listed in Table 3.…”

Bagged Tree And ResNet Based Joint End-to-End Fast CTU Partition Decision Algorithm For Video Intra Coding

“…These feature candidates are from four fields, which are information from neighboring CTUs, side information during pre-encoding process, statistical information of CTU pixels and information from pixel filtering results. We will give a brief introduction of these feature candidates in the following paragraphs, and more details are available in our previous work [32]. Results of [33] suggested that information from neighboring CTUs is useful for the partition decision making of current CU.…”

“…Many existing works have developed numbers of useful attributes for CU splitting label prediction. Based on these existing features, in our previous work [32] we also designed several novel features proven to be effective. As a result, based on our particular task of this paper we select totally 29 feature candidates used in [32], and they are listed in Table I along with corresponding meanings.…”

“…Based on these existing features, in our previous work [32] we also designed several novel features proven to be effective. As a result, based on our particular task of this paper we select totally 29 feature candidates used in [32], and they are listed in Table I along with corresponding meanings. These feature candidates are from four fields, which are information from neighboring CTUs, side information during pre-encoding process, statistical information of CTU pixels and information from pixel filtering results.…”

“…Specifically, the bagged tree is used to make the splitting decision for a 64×64 sized CU, while the ResNet is used to [36] as training sequences, which are Traffic (2560×2600), ParkScene (1920×1080), BasketballDrill (832×480), BQSquare (416×240) and FourPeople (1280×720). Then, these five sequences are encoded by HM16.7 under all intra configurations but four QPs (22,27,32,37).…”