Scholarlink

<p>The increased demand for compression of high definition videos with high efficiency leads to the development of High Efficiency Video Coding (HEVC) standards. The HEVC compresses video sequences with 50% less bit-rate compared to the H.264 standard. However, high efficiency is achieved at the cost of high encoding time. Most of the encoding time is consumed during the Rate-Distortion Optimization (RDO) search process and motion estimation. During the RDO search process, inter prediction is performed, which utilizes the Test Zone (TZ) search algorithm for motion estimation. This paper proposed a deep learning approach that predicts the Coding Unit (CU) size using a Convolutional Neural Network (CNN) rather than the conventional RDO search process. The Multi-Resolution frame with the Cross Diamond Octagonal search pattern (MRCDO) is proposed, which reduces the time required for motion estimation. The combination of the CU size prediction and the MRCDO search method reduce the encoding time by 66.91%, superior to state-of-the-art methods. The results also show that the RD performance loss of 9.16% rise in bit rate and 0.52dB loss in video quality on average, which is slightly higher compared to state-of-the-art methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fast Convolutional Neural Network based coding unit size prediction in HEVC

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Karthik Sairam Sanagavarapu,
  • Muralidhar Pullakandam

摘要

The increased demand for compression of high definition videos with high efficiency leads to the development of High Efficiency Video Coding (HEVC) standards. The HEVC compresses video sequences with 50% less bit-rate compared to the H.264 standard. However, high efficiency is achieved at the cost of high encoding time. Most of the encoding time is consumed during the Rate-Distortion Optimization (RDO) search process and motion estimation. During the RDO search process, inter prediction is performed, which utilizes the Test Zone (TZ) search algorithm for motion estimation. This paper proposed a deep learning approach that predicts the Coding Unit (CU) size using a Convolutional Neural Network (CNN) rather than the conventional RDO search process. The Multi-Resolution frame with the Cross Diamond Octagonal search pattern (MRCDO) is proposed, which reduces the time required for motion estimation. The combination of the CU size prediction and the MRCDO search method reduce the encoding time by 66.91%, superior to state-of-the-art methods. The results also show that the RD performance loss of 9.16% rise in bit rate and 0.52dB loss in video quality on average, which is slightly higher compared to state-of-the-art methods.