MRA-IDN: A Lightweight Super-Resolution Framework of Remote Sensing Images Based on Multiscale Residual Attention Fusion Mechanism

The emergence of deep-learning technology has significantly improved the performance of super-resolution algorithms for a single remote sensing image; however, the number of deep-learning model parameters is large, which limits its real-time deployment. In addition, the reconstructed image quality still needs improvement. To deal with the aforementioned problems, a lightweight multiscale residual attention information distillation network is proposed in this article. It achieves high-quality and fast super-resolution processing of remote sensing images. First, a high- and low-frequency (HF and LF) separation reconstruction strategy is adopted that enables the network to improve the reconstructed details of HF components while keeping the number of model parameters low. Second, a novel multiscale residual attention information distillation group is designed as the key component to further extract richer regional features with different perceptual fields and HF information while reducing the number of network parameters. This is achieved by combining a multiscale residual information distillation block that consists of multiple residual convolutional sub-blocks and an HF channel aware attention block. Last, the experimental results show that, compared with existing mainstream methods, such as the MHAN, the number of model parameters can be reduced by 75%, and the edge details of the reconstructed images are richer and more complete. The corresponding peak signal-to-noise ratio and SSIM can reach 31.59 dB and 0.824, respectively, under the condition of a
$ \times 4$
magnification factor, and 27.39 dB and 0.668, respectively, under the condition of a
$ \times 8$
magnification factor.