Super Resolution of Aerial Images of Intelligent Aircraft via Multi-scale Residual Attention and Distillation Network

Nowadays, aerial images of intelligent aircraft are widely used in all aspects of production and life. However, due to the limitations of airborne equipment, aerial images often have problems of low precision and high noise. Although, super resolution (SR) based on convolution neural network (CNN)can solve the above problems to some extent, huge number of parameters and computational overhead make these algorithms difficult to deploy on onboard computers. To address such limitations, a multi-scale residual attention and distillation network (MRADN) is designed. Firstly, with the aim to maintain the network lightweight enough, a multi-scale distillation network using depth-wise separable convolution (DSC) is proposed. Then, a multi-scale residual channel attention block (MS-RCAB) is designed, which leads to the network pays more attention to high-frequency details. What’s more, for the purpose of using feature information from different scales, a multi-scale attentional feature distillation block (MS-AFDB) is constructed. Compared with the networks in recent years, MRADN has advantages in parameters, computational complexity, processing speed and accuracy which has been proved by large number of experiments. Furthermore, experiments on self-built dataset has determined that this network is suitable for aerial images.