CMSnet: State of the Art Deep Learning Multiscale Reconstruction for Compressive Sensing

Abstract

We present what is, to the best of our knowledge, state-of-the-art reconstruction results for deep learning-based multiscale compressive sensing. Our reconstruction method is compared to a variety of recent compressive sensing reconstruction methods.

Comparison of deep learning-based compressive imaging from a practitioner's viewpoint

“…The upper row shows the CI results at a compression ratio 𝐶𝑅 = 𝑀 𝑁 of 10% obtained with AMP-Net, OPINE-Net and CMS-Net in the noiseless case. CMS-Net outperforms other methods, which corroborates the comparison presented in[15]. The lower row in Figure2demonstrates CI reconstruction at same CR=10%, but with measurement SNR of 20dB.…”

“…SSIM -This metric was developed to create quality measure which compatible with the human eye by measuring the similarity between images [14]. [15]. The lower row in Figure 2 demonstrates CI reconstruction at same CR=10%, but with measurement SNR of 20dB.…”