WDU-Net: Wavelet-Guided Deep Unfolding Network for Image Compressed Sensing Reconstruction

More and more deep unfolding networks have been studied to obtain good interpretability for high-quality image Compressed Sensing (CS) reconstruction. However, most of these networks simply focus on transmitting information across adjacent stages in the image-domain and ignore that frequency-domain information also can greatly restrain the solutions of CS optimization model. Motivated by this observation, frequency-domain consistency constraint is proposed to be inserted into image CS reconstruction. Concretely, we build a new CS optimization model based on discrete wavelet transform, in which frequency-domain information is used to guide image CS reconstruction. This model is divided into two sub-problems, which are optimized in an iterative manner. The iterative optimization procedure is expanded as a Wavelet-guided Deep Unfolding Network (WDU-Net) for image CS reconstruction. In view of the fact that denoising is the key step of the CS reconstruction problem, a dual-domain guided filtering block and a self-guided filtering enhancement block are proposed to remove noises for image reconstruction. Experimental results have shown that the reconstruction performance of our method is beyond many explainable CS reconstruction methods.