Compressed Remote Sensing by Using Deep Learning

Abstract

In remote sensing applications, storing and compressing images requires high memory size and consumes a lot of battery power, which can be overcome with compressed sensing (CS) and transferring complexities to the receiver. But compressed sensing has three significant challenges. At first, it is hard to find the basis in which signal is sparse. Second, compressed sensing uses recovery algorithms that are slow in time, which makes CS suitable for applications that are non-real-time. Third, CS usually use pre-specified measurement matrices, which are not optimized based on the data being analyzed, so it is possible to improve the performance of CS. In this paper, we will show that using deep learning methods in compressed remote sensing, the above challenges will be solved. In this paper, we present a deep neural network in which the matrix of measurements and reconstruction operations are optimized simultaneously. We indicate that this network closely approximates the recovery algorithms and has performance near CS recovery algorithms, but it is faster in run time. However, the complexity of deep neural networks is the training phase of the network and needs to be completed only once before using the network.