关键词: 视频压缩感知, 神经网络, 时域注意力, 特征对齐, 运动补偿, 深度重构

Abstract:

The motion compensation methods of optical flow alignment and deformable convolution alignment adopted by the existing video compressed sensing reconstruction algorithms have problems such as error accumulation and limited information perception range, which greatly limit their effectiveness and practicability. In order to adaptively extract the global information of the reference frame without introducing extra parameters, this paper first proposes an innovative idea of using the attention mechanism to realize motion estimation and motion compensation in video compressed sensing reconstruction, and then designs the temporal-attention feature alignment network (TAFA-Net) for implementation. On this basis, a joint deep reconstruction network (JDR-TAFA-Net) is proposed to achieve high-performance reconstruction for non-key frames. First, the reference frames are adaptively aligned to the current non-key frame through TAFA-Net, and then a fusion network based on the auto-encoder is introduced to fully extract the relevant information from existing frames to further enhance the reconstruction quality of the non-key frames. Experimental results show that, compared with the state-of-the-art iterative optimization-based method SSIM-InterF-GSR, the proposed method can improve PSNR (Peak Signal to Noise Ratio) by 4.74dB, and compared with the state-of-the-art deep learning-based method STM-Net, the proposed method can improve PSNR by 0.64dB.

Key words: video compressed sensing, neural network, temporal attention, feature alignment, motion compensation, deep reconstruction