As a crucial part of video compression, intra prediction utilizes local information of images to eliminate the redundancy in spatial domain. In both H.265/HEVC and H.266/VVC, multiple directional prediction modes are employed to find the texture trend of each small block and then the prediction is made based on reference samples in the selected direction. Recently, the intra prediction schemes based on neural networks have achieved great success. In these methods, the networks are trained and applied to intra prediction in addition to the directional prediction modes. In this paper, we propose a novel data clustering-driven neural network (dubbed DCDNN) for intra prediction, which can learn deep features of the clustered data. In DCDNN, each network can be split into two networks by adding or subtracting Gaussian random noise. Then a data clustering-driven training is applied to train all the derived networks recursively. In each iteration, the entire training dataset is partitioned according to the recovery qualities of the derived networks. For the experiment, DCDNN is implemented into HEVC reference software HM-16.9. The experimental results demonstrate that DCDNN can reach an average of 4.2% Bjontegaard distortion rate (BDrate) improvement (up to 7.0%) over HEVC with all intra configuration. Compared with existing fully connected networkbased intra prediction methods, the bitrate saving performance is further improved.

中文翻译：

---

用于帧内预测的数据聚类驱动的神经网络

作为视频压缩的关键部分，帧内预测利用图像的局部信息来消除空间域的冗余。在H.265/HEVC和H.266/VVC中，均采用多种方向预测模式来发现每个小块的纹理趋势，然后根据所选方向的参考样本进行预测。最近，基于神经网络的帧内预测方案取得了巨大的成功。在这些方法中，除了方向预测模式之外，网络还经过训练并应用于帧内预测。在本文中，我们提出了一种用于帧内预测的新型数据聚类驱动神经网络（称为 DCDNN），它可以学习聚类数据的深层特征。在 DCDNN 中，可以通过添加或减去高斯随机噪声将每个网络拆分为两个网络。然后应用数据聚类驱动的训练来递归地训练所有派生网络。在每次迭代中，整个训练数据集根据派生网络的恢复质量进行分区。在实验中，DCDNN 在 HEVC 参考软件 HM-16.9 中实现。实验结果表明，在所有帧内配置下，DCDNN 可以达到比 HEVC 平均 4.2% 的 Bjontegaard 失真率 (BDrate) 改进（高达 7.0%）。与现有的基于全连接网络的帧内预测方法相比，进一步提高了比特率节省性能。DCDNN 在 HEVC 参考软件 HM-16.9 中实现。实验结果表明，在所有帧内配置下，DCDNN 可以达到比 HEVC 平均 4.2% 的 Bjontegaard 失真率 (BDrate) 改进（高达 7.0%）。与现有的基于全连接网络的帧内预测方法相比，进一步提高了比特率节省性能。DCDNN 在 HEVC 参考软件 HM-16.9 中实现。实验结果表明，在所有帧内配置下，DCDNN 可以达到比 HEVC 平均 4.2% 的 Bjontegaard 失真率 (BDrate) 改进（高达 7.0%）。与现有的基于全连接网络的帧内预测方法相比，进一步提高了比特率节省性能。