Reconstructing images using compressive sensing (CS) for real-time applications demands less processing, and better runtime. With the advent in artificial intelligence (AI) tools image compressive sensing is feasible. Among the recent techniques, deep learning approaches have emerged as crucial in various fields such as computer vision, remote sensing, image processing, and signal processing due to their exceptional performance in recovering images compared to traditional methods. In today’s big data era, there is a big demand for high quality photographs that can be retrieved from low-quality images using the principles of CS. One such method, known as pixel shuffle, is used within the framework of deep learning and convolutional neural networks (CNN) to enhance the spatial resolution of images. This study implements the unsupervised learning in a deep network constructed with unfold networks to realise Chamboulle-Pock algorithm along with pixel shuffle techniques. The network is trained using the Berkeley Segmentation Dataset 500 (BSD500) with a peak signal-to-noise ratio (PSNR) of 35 dB. The PSNR and structural similarity index measure (SSIM) of the reconstructed image in the proposed method dubbed as CSCP Net is compared with those of other CS deep networks, including CSNet+, CSNet#, TVAL, ISTANet, ISTANet++.
Sukumaran et al. (Thu,) studied this question.