MFFAGAN: Generative Adversarial Network With Multilevel Feature Fusion Attention Mechanism for Remote Sensing Image Super-Resolution

Abstract

Super-resolution (SR) based on deep learning has been playing an important role in improving the spatial resolution of remote sensing images. Although convolutional neural networks (CNNs) dominate the research of remote sensing image SR, most of them struggle to fully utilize the multilevel features in information transmission. Constantly expanding the network architecture sometimes also leads to an increase in feature redundancy and computational complexity. Moreover, CNN-based methods are unable to generate visually appealing images. To address the aforementioned issues, we propose a multilevel feature fusion attention SR method based on GAN called MFFAGAN. Specifically, we propose a novel enhanced mixed-attention block (EMAB), which enables the network to capture key feature information in both the channel and spatial domains. Meanwhile, in order to enhance the model's ability to extract various features at multiple levels more efficiently, we propose a multilevel feature fusion attention module (MFFAM). The output of each residual block is directly fed into the feature aggregation block and eventually combined with the attention branch. Thus, the network is capable of aggregating these information-rich residual features without any loss to produce more representative features. Experimental results show that our proposed MFFAGAN outperforms most state-of-the-art methods in both quantitative and qualitative metrics.