Structure Prior Guided Deep Network for Compressive Sensing Image Reconstruction From Big Data

Recent research on image reconstruction from Compressive Sensing theory has achieved great success due to the devolopment of deep convolutional neural networks(DCNN). However, commonly used loss function suffers from a limitation in that the reconstruction data should obey an unimodal Gaussian distribution which is disobedience to the real distribution of natural image data [1]. In this work, we propose an structure prior guided compressive sensing network called SGCSNet. The key portion of SGCSNet is the adaptive structure guided block(dubbed as ASGB) in reconstruction network. By fully using the image structure prior as guidance constraint to decompose the multi modal distribution into multiple unimodal distributions based on ASGB, SGCSNet could improve the visual effect of the details in restored image for commonly used loss function like $l_{2}$ loss. In addition, we introduce gram matrix instead of $l_{2}$ norm to depict the distance between feature layers. The gram matrix can express the difference of features from a global views and improve the texture details of image restoration. Experiments show that SGCSNet has a better performance and could obtain more realistic results than the existing state of the art method.