Adaptive deep learning network for image reconstruction of compressed sensing

In this paper, we study how to achieve sparse sampling and high-quality reconstruction of natural images, and propose an interpretable deep network based on proximal gradient descent (PGD), dubbed AICS-Net, while performing joint constraint optimization of adaptive sparse sampling and reconstruction of images. AICS-Net consists of a sampling sub-network, an initialization sub-network and a recovery sub-network. The sampling sub-network achieves adaptive sampling of images, and the initialization sub-network uses the transpose of the measurement matrix to achieve initialized reconstruction of images. Integrating the gradient estimation strategy into the gradient descent step of the PGD algorithm in the recovery sub-network, then an optimization-based staged network structure is constructed. Moreover, to increase the hardware reputability of AICS-Net, binary constraint and orthogonal constraint are added to the measurement matrix. To improve the quality and visual effect of reconstructed images, a loss function is created to account for the color and texture differences between images. Experimental results show that the proposed AICS-Net can achieve better image reconstruction while maintaining reconstruction speed compared to existing state-of-the-art network-based CS algorithms, especially at low CS ratios. When the CS ratio is 1%, for Set11, PNSR and SSIM can be improved by at least 6.63% and 7.57%, respectively.