Search Articles

Home / Articles

Image Compressed Sensing Using Deep Learning

. Saksham Gera, Shiva Gupta, Paras HS, Rohan Pillai and Madan Mohan Rayguru


Abstract

In this paper, techniques used to compress data, especially images are discussed. A deep learning model based on convolution neural networks is implemented and the results in the form of PSNR (Peak Signal to Noise Ratio) are compared. The main hurdles in compressed sensing of images are the sampling matrix design and the reconstruction method. To overcome these we propose a CNN framework called CS Net, which consists of a joined sampling network and reconstruction network. The sampling network adapts the sampling matrix during the training of data, which enables the compressed sensing measurements to maintain additional structural data for better recreation of images. The reconstruction network is divided into two parts, an elementary linear network, and a deep reconstruction non-linear network; it adapts a mapping between the endpoints of structural measurements and reconstructed images. In this project, we used a different number of blocks in the deep reconstruction and compared the results. The PSNR values of the reconstructed images gradually increase as we increase the number of residual blocks, but after a certain point, the values start decreasing[1].

The results obtained show that CSNet produces high-quality images like those obtained from methods like FFT and Nyquist Shanon but with comparatively low computation costs[2][3].

KeywordsCompressed Sensing, Convolution Neural Networks, CSNet, Sampling matrix, Image reconstruction.

Download :