基于时空特征的视频压缩自适应缩放方法  

Adaptive Scaling Method Based on Spatial-Temporal Feature for Video Compression

作　　者：吴佩颖 沈礼权[1] 

机构地区：[1]上海大学通信与信息工程学院,上海200444

出　　处：《工业控制计算机》2025年第2期102-104,共3页Industrial Control Computer

基　　金：国家自然科学基金(61931022)。

摘　　要：深度学习技术越来越多地集成到视频压缩框架中,并显著提高了压缩效率。然而,现有方法受限于两点:一是忽略了重缩放过程中不同特征冗余间的相互作用,难以有效促进压缩;二是当前方法通常对所有类型的帧采用统一缩放方法,导致关键信息丢失。为了解决上述问题,提出一种基于时空特征的重缩放框架,该框架包括时空去冗余网络和自适应缩放网络。时空去冗余网络根据帧的特性动态选择去冗余方法,消除关键帧的空域冗余和非关键帧的时域冗余,并保留了关键细节,为后续的帧预测提供了丰富的信息。自适应缩放网络实现输入输出双向映射,确保缩放和编码过程中信息的连续性和完整性。实验结果显示,与编解码器锚点相比,该框架在低时延配置下实现15.73%的BDBR降低,证明了所提算法的有效性。This paper proposes a rescaling framework based on spatial-temporal features,which consists of a spatial-temporal redundancy removal network and an adaptive scaling network.The spatial-temporal redundancy removal network dynamically selects redundancy removal strategies based on frame characteristics,effectively eliminating spatial redundancy in key frames and temporal redundancy in non-key frames.It also preserves detailed information in key frames to provide a reliable basis for subsequent frame prediction.The adaptive scaling network ensures the continuity and integrity of information during scaling and encoding processes by implementing bidirectional mapping between input and output.Extensive experiments show that our network achieves a significant 15.73%BDBR reduction under low delay P configuration compared to the codec anchor,demonstrating the effectiveness of the proposed method.

关 键 词：视频压缩 时空冗余 视频缩放 

分 类 号：TN919.81[电子电信—通信与信息系统] TP18[电子电信—信息与通信工程]