An Efficient Distributed Video Compressed Sensing Method

In recent years, various methods for deep learning techniques for image compressed sensing (CS) have seen rapid advancements, achieving high-quality image reconstruction while maintaining low computational costs. Nevertheless, current deep learning-based image CS techniques primarily concentrate on correlations within frames and often overlook correlations between frames, which leads to inefficiencies when applied directly to video compressed sensing. In this paper, we introduce an effective method for distributed video compressed sensing that aims to leverage both intra- and inter-frame correlations. The proposed distributed video compressed sensing system integrates the benefits of CS with distributed video coding, making it suitable for environments with finite resources for video compression and transmission. However, traditional compressed video sensing (CVS) methods are computationally intensive and time consuming, limiting their real-time applications. We propose a fast CVS reconstruction method based on deep network, named FMCDR-Net. We first perform ISTA-Net<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">++</sup> as the initial reconstruction. The inter-frame reconstruction consists of two stages, where the first stage obtains the initial enhanced reconstructed non-key frames by the designed FMC-Net for motion compensation, and the second stage obtains the final reconstructed non-key frames by the designed FDR-Net for residual reconstruction using the frame difference information in order to recover the details. Experiments manifest that our network performs favorably against similar methods with the better trade-off between parameters and performance.