LAUSR

Single image super resolution (SISR) based on deep learning has made great progress in recent years. As the method continues to improve, different network structures have been proposed to better perform SR feature extraction for reconstruction. A deep structure has a good ability to generate high-quality SR features, but the complex structure may also cause problems such as hard training and overfitting. Many efforts have also been made to solve these problems, such as feedback structure and attention mechanism. However, naively applying these methods to SR networks may be useless. Hence, in this research, we took a further step by introducing progressive residual attention to generate high-quality SR images. In experiments, we compared the reconstruction results and training progress with other SR methods based on normal structures. The proposed network achieves fast convergence speed and better SR results.