Deep Learning methods for Compressed Sensing image reconstruction

Deep Learning methods for Compressed Sensing image reconstruction Deep learning (DL) has become, in the last years, the principal way to process images in all the fields of image elaboration and the same goes for Compressive Sensing (CS), the new technique that is making its way in the world of image processing. Compressive Sensing is a full of potential technique, above all for what concerns the acquisition of images, in the case of MRI for example, or the transmission of very big images, e.g. the transmission of the images from satellites. In fact, CS, allowing, together with the single-pixel camera, to directly acquire very compressed data, avoid a waste of time and computational resources to acquire and compress images. Traditional compression techniques are not universal techniques and they suit up for the data, understanding each time how to compress it in the better way. CS, instead, provides a universal way for data acquisition that will speed up these processes and will allows to better and more easily transmit these data too. Unfortunately, CS is not so good in reconstructing the acquired signals and this is its biggest problem. In fact, the iterative classical methods, like minimization of l1-norm, could also take days to exactly reconstruct an image and all the gain of the acquisition and transmission phase is lost. DL aims at solving this problem and, even if the accuracy could not be of 100% of course, the latest results show its great potential. In this thesis, we have evaluated the efficiency of these DL based methods for CS reconstruction and developed an algorithm considering various block size (an important parameter in image processing with NN), fixed and trained sensing matrices and various CS ratios. The results obtained far exceed the state of the art in terms of accuracy and PSNR and remain comparable in terms of elaboration time. Moreover, the algorithm developed works very well with every kind of image.