High Frequency Feature Distillation Network for Compressive Sensing Reconstruction

Deep unfolding network for reconstructing compressive sensing images has achieved remarkable success in recent years owing to its superior interpretability and performance. Nonetheless, the deficiency of high-frequency components in the reconstructed images remains a critical issue. Moreover, the utilization of a deep and wide network leads to a computational burden, which poses a significant challenge. In this paper, High Frequency Feature Distillation Network (HFFDNet) is proposed to solve these issues. In HFFDNet, a High Frequency Enhancement Unit (HFEU) is employed to extract and enhance the high frequency components. To further refine it with fewer parameters, we design a High Frequency Feature Distillation Block (HFFDB) that learns only a portion of the feature channels obtained by HFEU and distills the rest without learning. Our experimental results demonstrate that HFFDNet is a promising approach for image CS reconstruction, which saves the number of parameters up to 96% with better quality compared to state-of-the-art methods.