Abstract
In the era of the Internet of Things (IoT), the rapid growth of interconnected devices has intensified the demand for efficient data acquisition and processing techniques. Compressive Sensing (CS) has emerged as a promising approach for simultaneous signal acquisition and dimensionality reduction, particularly in multimedia applications. In response to the challenges presented by traditional CS reconstruction methods, such as boundary artifacts and limited robustness, we propose a novel hierarchical deep learning framework, SwinTCS, for CS-aware image reconstruction. Leveraging the Swin Transformer architecture, SwinTCS integrates a hierarchical feature representation strategy to enhance global contextual modeling while maintaining computational efficiency. Moreover, to better capture local features of images, we introduce an auxiliary convolutional neural network (CNN). Additionally, for suppressing noise and improving reconstruction quality in high-compression scenarios, we incorporate a Non-Local Means Denoising module. The experimental results on multiple public benchmark datasets indicate that SwinTCS surpasses State-of-the-Art (SOTA) methods across various evaluation metrics, thereby confirming its superior performance.