Perceptual Compressive Sensing

This paper proposes perceptual compressive sensing. The network is composed of a fully convolutional measurement and reconstruction network. For the following contributions, the proposed framework is a breakthrough work. Firstly, the fully-convolutional network measures the full image which preserves structure information of the image and removes the block effect. Secondly, with the employment of perceptual loss, we no longer concentrate on the Euclidean distance of reconstruction. Instead, we just care about the perceptual invariance, which makes the reconstructed image obtain much semantic information. We call it semantics-oriented reconstruction. Experimental results show that the proposed framework outperforms the existing state-of-the-art methods.