FSMANet: Feature-enhanced Sampling and Multi-scale Attention Reconstruction Network for Image Compressive Sensing

As vital carriers of information, images require efficient compression methods, among which compressive sensing (CS) has emerged as a prominent technique. Deep Unfolding Networks have significantly advanced the field of CS-based image reconstruction by integrating traditional optimization algorithms with deep learning, thereby improving both performance and interpretability. However, existing methods often suffer from two major limitations: suboptimal design in sampling mechanisms and insufficient optimization of multi-scale features during reconstruction, which ultimately restrict their ability to recover fine details. To address these issues and enhance multi-scale representation within deep CS frameworks, we propose a novel Feature-enhanced Sampling and Multi-scale Attention Reconstruction Network (FSMANet). In the sampling subnet, we introduce a learnable feature-enhanced sampling module that facilitates end-to-end collaborative learning between sampling and reconstruction. Driven by the reconstruction objective, this subnet actively learns to capture measurements that best preserve multi-scale detail information in the image. Within the reconstruction subnet, we design a dual-path feature denoising module that combines convolutional networks with attention mechanisms to adaptively and discriminatively extract image features across different scales. Finally, we conduct comprehensive comparative experiments between the proposed FSMANet and existing state-of-the-art methods. Analytical results demonstrate that FSMANet consistently outperforms current leading approaches across multiple public datasets.