基于纹理结构与CA-Transformer的深度图帧内快速编码算法OA
A fast intra-frame coding algorithm for depth maps based on texture structure and CA-Transformer
针对三维高效视频编码(three-dimensional high efficiency video coding,3D-HEVC)在深度图编码中复杂度过高的问题,本文提出了一种基于纹理划分结构与深度学习并行的算法实现深度图的快速编码算法.首先,提出了基于纹理图划分结构加速深度图编码的方法,提前识别并简化编码树单元(coding tree unit,CTU),得到快速划分结果.其次,针对无法使用纹理图划分结构快速划分的 CTU,设计了一种名为 CA-Transformer 的神经网络.应用卷积神经网络对局部特征的敏感性与视觉 Transformer 对全局特征的敏感性,将结构复杂的深度图图像输入到本文的 CA-Transformer 网络中,得到对深度图 CTU划分结构的预测.使用本文算法替换原始编码标准的 CTU 划分过程,实验结果表明,本文算法在保证率失真性能与合成视点质量几乎不受影响的前提下,平均减少了 56.00%的编码时间.
To address the high complexity issue in depth map coding of three-dimensional high efficiency video coding(3D-HEVC),this paper proposes a fast depth map coding algorithm that integrates texture partitioning structure and deep learning.First,a texture map partitioning structure based method is introduced to accelerate depth map cod-ing,which identifies and simplifies coding tree unit(CTU)in advance to achieve rapid partitioning results.Sec-ond,a neural network named CA-Transformer is designed.Leveraging the sensitivity of convolutional neural net-works to local features and the global feature modeling capability of vision Transformers,the proposed CA-Trans-former network processes structurally complex depth map images to predict CTU partitioning structures.By repla-cing the original CTU partitioning process in the standard encoder with the proposed algorithm,experimental results demonstrate that the algorithm reduces encoding time by an average of 56.00%while maintaining nearly identical rate-distortion performance and synthesized view quality.
李竟泽;贾克斌
北京工业大学信息科学技术学院 北京 100124先进信息网络北京实验室 北京 100124
三维高效视频编码纹理图深度图编码树单元划分帧内编码
three-dimensional high efficiency video codingtexture mapdepth mapcoding tree unit parti-tioningintra encoding
《高技术通讯》 2026 (7)
713-723,11
北京市自然科学基金(4212001)资助项目.
评论