基于纹理特征的分布式视频压缩感知自适应重构方法

Adaptive Reconstruction for Distributed Compressive Video Sensing Based on Texture Features

面对大规模视频数据带来的全新挑战,具备硬件友好特性的分布式视频压缩感知应运而生。由于传统基于分析模型的分布式视频压缩感知重构方法计算复杂度高,难以满足实时应用的要求,因此深度学习技术被逐渐引入。然而,现有基于深度学习的重构方法忽略了帧的纹理特征,限制了重构性能。由于同图像组中的视频帧具有较高的相似性,因此可以选择相邻视频帧作为当前视频帧纹理特征的参考。为了解决这个问题,提出一种基于纹理特征的分布式视频压缩感知自适应重构网络,命名为TF-DCVSNet。具体来说,TF-DCVSNet利用已重构的相邻帧纹理特征,激活当前重构帧的重构网络模块,进行自适应重构。大量实验验证了TF-DCVSNet的有效性。

Distributed compressive video sensing(DCVS),possessing hardware-friendly characteristics,has emerged as a solution to the challenges posed by large-scale video data.Since traditional analytical model-based reconstruction methods for DCVS are computational-ly complex and hard to meet the requirements of real-time applications,deep learning technologies are gradually applied in DCVS.How-ever,the existing deep learning-based reconstruction methods ignore texture characterizes of frames,which limits reconstruction perform-ance.Based on the observation that frames within one group are highly correlated,the adjacent frames can be selected as the references to exploit texture features of current frames.To address this problem,a texture features-based adaptive reconstruction network for DCVS is proposed,dubbed‘TF-DCVSNet’.Specifically,TF-DCVSNet utilizes the texture features of the reconstructed adjacent frames to acti-vate the reconstruction network module of the current reconstructed frame to perform adaptive reconstruction.Extensive experiments demonstrate the effectiveness of TF-DCVSNet.