Visualizing and understanding of learned compressive sensing with residual network

摘要

Recent years a variety of CNN-based (convolutional neural network) approaches for compressive sensing (CS) have been proposed. They learn a transform to recover the original signals from the measurements obtained by measuring the scene at a sub-Nyquist sampling rate. Among them, the LMM-based ones (learned measurement matrix) exhibit better performance. In this paper, we visualize the LMM-based CS framework. This is the first time an insight look is taken into the CS network. It helps us understand how CS framework works. Taking the proposed LMM-based framework as an example, where reasonable residual blocks in the recovery part let it achieve excellent performance over the existing ones, we analyze the mechanism of CNN-based CS by the visualization. In the measurement part, intuitive representation of the measurement matrices is presented. As for the recovery procedure, an explanation of the preliminary recovery is given from the viewpoints of system and space. We analyze how the residual block adds the mainly high-frequency information. Through the comparison of the visualization of the typical methods, it is explored that the measurement and recovery part of the proposed method can promote each other, and the learned CS framework with residual network achieves the best performance. In addition, a set of experiments are conducted on a standard dataset to verify the better performance of our framework. (C) 2019 Elsevier B.V. All rights reserved.