Non-Local Prior Dense Feature Distillation Network for Image Compressive Sensing

Deep learning-based image compressive sensing (CS) methods often suffer from high computational complexity and significant loss of image details in reconstructions. A non-local prior dense feature distillation network (NPDFD-Net) is proposed for image CS. First, the non-local priors of images are leveraged to enhance high-frequency information in the measurements. Second, a discrete wavelet decomposition learning module and an inverse discrete wavelet reconstruction module are designed to reduce information loss and significantly lower computational complexity. Third, a feature distillation mechanism is incorporated into residual dense blocks to improve feature transmission efficiency. Finally, a multi-scale enhanced spatial attention module is proposed to strengthen feature diversity. Experimental results indicate that compared to MRCS_GAN, OCTUF, and DPC-DUN, the proposed method achieves an average PSNR improvement of 1.52%, 2.35%, and 0.93%, respectively, on the Set5 dataset. The image reconstruction running time is enhanced by 93.93%, 71.76%, and 40.74%, respectively. Furthermore, the proposed method exhibits significant advantages in restoring fine texture details in the reconstructed images.