CNN-based in-loop filtering for coding efficiency improvement
CNN-based in-loop filtering for coding efficiency improvement
复制标题
DOI：
10.1109/ivmspw.2016.7528223
复制
发表时间：
2016-07
期刊：
影响因子：
--
通讯作者：
Woonsung Park;Munchurl Kim
中科院分区：
文献类型：
--
作者：
Woonsung Park;Munchurl Kim
最近的视频编码标准称为高效视频编码（HEVC），采用两个环路滤波器来提高编码效率，其中环路滤波由去块滤波器（DF）和样本自适应偏移（SAO）滤波完成。 DF 有助于提高编码效率和主观质量，而无需向解码器侧发送任何比特信号，而 SAO 滤波则通过向解码器发送偏移值来纠正量化误差。在本文中，我们首先提出了一种使用卷积神经网络（CNN）的新环路滤波技术（称为 IFCNN），用于提高编码效率和主观视觉质量。通过在编码器和解码器中使用相同的训练权重，IFCNN 不需要信令位。所提出的 IFCNN 在两个不同的 QP 范围内进行训练：QR1 从 QP = 20 到 QP = 29； QR2从QP = 30到QP = 39。在测试中，在QR1中训练的IFCNN应用于QP值小于30的编码/解码，而在QR2中训练的IFCNN应用于QP值大于29的情况。实验结果表明，所提出的IFCNN优于HEVC参考模式（HM），低延迟的BD速率平均增益为1.9％-2.8％配置中，IDR 周期为 16 的随机接入配置的 BD 速率平均增益为 1.6%-2.6%。
A recent video coding standard, called High Efficiency Video Coding (HEVC), adopts two in-loop filters for coding efficiency improvement where the in-loop filtering is done by a de-blocking filter (DF) followed by sample adaptive offset (SAO) filtering. The DF helps improve both coding efficiency and subjective quality without signaling any bit to decoder sides while SAO filtering corrects the quantization errors by sending offset values to decoders. In this paper, we first present a new in-loop filtering technique using convolutional neural networks (CNN), called IFCNN, for coding efficiency and subjective visual quality improvement. The IFCNN does not require signaling bits by using the same trained weights in both encoders and decoder. The proposed IFCNN is trained in two different QP ranges: QR1 from QP = 20 to QP = 29; and QR2 from QP = 30 to QP = 39. In testing, the IFCNN trained in QR1 is applied for the encoding/decoding with QP values less than 30 while the IFCNN trained in QR2 is applied for the case of QP values greater than 29. The experiment results show that the proposed IFCNN outperforms the HEVC reference mode (HM) with average 1.9%-2.8% gain in BD-rate for Low Delay configuration, and average 1.6%-2.6% gain in BD-rate for Random Access configuration with IDR period 16.