ISTOF-Net: An ISTA-Based Deep Unfolding Network With Optimized Feature Aggregation Architecture for Image Compressed Sensing

Compressed sensing (CS) has emerged as an efficient framework for sensor-based data acquisition and reconstruction, and is now widely employed in image processing. Deep unfolding networks (DUNs) integrate deep neural networks with optimization algorithms, emerging as a prevalent approach for image CS by enhancing interpretability and reconstruction performance. However, most existing DUNs suffer from insufficient integration and utilization of feature information during the iterative process. To address this limitation and achieve the goal of fusing complementary information to form a richer and finer feature representation, we propose a iterative shrinkage-thresholding algorithm (ISTA)-based DUN with optimized feature aggregation architecture for high-quality image CS, dubbed ISTOF-Net. By introducing a feature aggregation architecture, ISTOF-Net enhances both the integration and utilization of local detail features and global structural features. Specifically, in the reconstruction module, we design a novel optimized feature aggregation module (OFAM), which combines global structural features with local detail features, enhancing their integration and utilization through feature modulation and attention mechanisms. Moreover, we introduce auxiliary convolution module (CM) to further refine and enhance fine image details. In addition, we adopt a lightweight network design, effectively reducing both computational complexity and reconstruction time. Experimental results on datasets acquired by image sensors and on remote sensing image demonstrate that ISTOF-Net outperforms state-of-the-art methods in both reconstruction quality and efficiency. Its high performance and lightweight design make ISTOF-Net ideal for sensor-based imaging systems that demand data compression and rapid imaging.