Abstract: With the development of deep learning, convolution-based image super-resolution has been widely applied in remote sensing. However, convolution is biased towards local features, making it difficult to extract global features. Transformer has the advantage of its self-attention mechanism for feature extraction of the global context. However, the high space and time complexity brings a non-negligible burden to the limited computing resources in remote sensing image processing. To solve the above issues, a remote sensing image super-resolution network that combines adaptive convolution and lightweight Transformer is proposed. An adaptive local feature extraction based on convolution is proposed to dynamically fuse fine-grained local features with different receptive field ranges in high and low frequency image features through adaptive weights. In addition, a lightweight Transformer structure is proposed, which decomposes the global self-attention mechanism into a local self-attention mechanism weighted by global information based on the task characteristics of remote sensing image super-resolution. The lightweight Transformer effectively reduces the computational complexity while maintaining the global perception capability. Experimental results show that the proposed method can better reconstruct the texture details of remote sensing images and achieve a good trade-off between computational complexity and super-resolution reconstruction quality.

Key words: image super-resolution, remote sensing image, adaptive convolution, lightweight Transformer

摘要： 随着深度学习的发展，基于卷积的图像超分辨在遥感领域取得了广泛应用，而卷积往往对局部特征更加敏感，难以提取遥感图像全局特征。Transformer因其自注意力机制能够对全局上下文进行特征提取而更具优势，但其高空间时间复杂度也给遥感图像处理有限的计算资源带来了不可忽视的负担。针对以上问题，提出一种自适应卷积与轻量化Transformer相结合的遥感图像超分辨网络。提出基于卷积的自适应局部特征提取，通过自适应权重动态融合高低频图像特征中不同感受野范围的细粒度局部特征。提出一种轻量化Transformer结构，结合遥感图像超分辨的任务特点将全局自注意力机制分解为全局信息引导下的局部自注意力机制，在保持全局感知能力的同时有效降低Transformer计算复杂度。实验结果表明，所提方法能够更好地重建遥感图像的纹理细节，实现计算复杂度和超分辨重建质量之间的良好权衡。

关键词: 图像超分辨, 遥感图像, 自适应卷积, 轻量化Transformer