Pyramid Attention Generative Adversarial Network for Image Compressed Sensing

The single image compressed sensing method has developed from the traditional model-based to the neural network such as generative adversarial network (GAN), which proves that the GAN-based method has better reconstruction quality and lower computational complexity. However, there are still difficulties in insufficient feature extraction in GAN that affect the quality of reconstruction. To solve the problem, this paper proposes an improved GAN-based CS model: Pyramid Attention-based GAN (PAGAN) for compressed sensing reconstruction with intra- & inter- attention mechanism. PAGAN reconstructs images from recurrent multiple sampled data and the proposed intra- & inter- attention-based generated network with different scale pyramid layers to enhance the multi-scale feature extraction and residual high-frequency feature transmission, which further improve the image resolution and the reconstruction quality and further enable CS reconstruction with flexible resolutions by combining multiple layers of the generated network. Experimental results on multiple public dataset show that PAGAN provides 3.12dB PSNR, and improves the average SSIM by 33.20%, compared with traditional and data-driven benchmark methods. It shows that the reconstructed image quality of the proposed method is superior to the previous methods.