Image Compressive Sensing via Multi-scale Feature Extraction and Attention Mechanism

In this paper, we proposed an end-to-end deep learning-based image compressive sensing framework. Our framework can adaptively sample original images and use linear and nonlinear transformation to reconstruct images. Multi-scale attention block (MAB), which adopts channel attention mechanism and multi-scale feature extraction, is applied in our network to extract meaningful features for recovering high frequency information. Deep supervision is also applied in the training stage to further enhance the reconstruction quality. Experimental simulation shows our framework has a competitive reconstruction performance compared with some classic methods.