CS-DeCNN: Deconvolutional Neural Network for Reconstructing Images from Compressively Sensed Measurements

Wentao Wan, Guohui Li, Peng Pan

Abstract: One important research point of compressive sensing (CS) is to restore a high-dimensional signal as completely as possible from its compressed form, which has much lower dimensionality than the original. Several methods have been employed to this end, including traditional iterative methods as well as recurrent approaches based on deep learning. This paper proposes a novel architecture of a deconvolutional neural network (CS-DeCNN) to address the signal reconstruction problem in CS. Compared with state-of-the-art CS methods, our proposed method not only has the advantage of high running speed; but also achieves higher recovery precision. Besides, CS-DeCNN perform better with less space occupancy compared to other deep learning CS methods. All these advantages make CS-DeCNN more suitable for embedded systems. We demonstrate the efficacy of our method by carefully performing comparative experiments.