Parallel multi-stream dual-aggregation unfolding paradigm for compressive sensing reconstruction

Deep unfolding architectures have advanced compressive sensing (CS) by combining model-driven and data-driven approaches, yet existing methods suffer from information loss and inefficient multi-view feature extraction. To address this, we propose PMD-Net, a Proximal Gradient Descent (PGD)-based network employing a Parallel Multi-Stream Iterative (PMSI) architecture that operates entirely in the high-dimensional feature domain. By embedding the iterative model within U-shaped cross-bridges to maintain high-throughput streams, PMD-Net maximizes multi-scale feature usage and prevents information loss caused by domain transformations. We further incorporate a Deep Gradient Fusion (DGF) unit for efficient gradient-guided information injection and a Dual Aggregated Information Fusion (DAIF) unit as the proximal operator. Integrating a Multi-Dconv Head Channel Transformer (MDCT) and a Shifted Window-Based Spatial Transformer (SWST), the DAIF unit enhances detail reconstruction by fusing multi-view features. Extensive experiments demonstrate that PMD-Net significantly outperforms state-of-the-art methods in natural image CS. Our code is available at https://github.com/nikou-arch/PMD-Net.