Sampling templates guided compression reconstruction network

For reconstruction in spatial compressive imaging, we use a module to fuse and extract the information in sampling templates, this obtained feature vector becomes the attention weight, which is multiplied with the feature maps of the compressed measurement frames. In addition, unlike previous networks using segmented images, we use full measurement frames collected as our network input. Thus the local information of objects can be preserved and blocky effect can be avoid. We have tested the network performance on the datasets, Set5, Set14, BSD100, Urban100, Manga109, with 25% compression rate, respectively. We obtain the PSNR\SSIM values in the range, [26.5dB,31.9dB]\[0.82,0.90]. This result is better than [23.6dB,29.0dB]\[0.72,0.85] obtained using the best algorithms in the same application based on our knowledge.