FEATURE-DOMAIN PROXIMAL HIGH-DIMENSIONAL GRADIENT DESCENT NETWORK FOR IMAGE COMPRESSED SENSING

Recently, by incorporating optimization theory into deep neural networks, optimization-inspired networks have achieved remarkable success in image compressed sensing (ICS). However, existing networks only learn a single gradient paradigm at each phase, making it difficult to fully utilize the measurements information. In this paper, we propose a novel idea of calculating gradients in high-dimensional space during the updating process to fully exploit measurement information. On this basis, we further propose the Feature-Domain Proximal High-dimensional Gradient Descent (FPHGD) algorithm to realize the proximal gradient descent in feature-domain and design a Feature-domain Proximal High-dimensional Gradient Descent network (FPHGD-Net) for ICS. Besides, to meet different requirements, we give three designs that combine different proximal mapping patterns in feature-domain to construct the networks. Extensive experiments demonstrate that reconstruction performance of our networks outperform existing state-of-the-art methods.