| 摘 要: 针对使用单一尺寸的卷积核重建遥感图像效果较差的问题,提出了融合多尺度信息和混合注意力的遥感图像重建模型。模型使用两种不同的多尺度特征提取块,能有效提取不同感受野下的特征图中的高频和低频特征,并通过混合注意力机制自适应地调整多尺度特征的权重,利用重建模块重建高清遥感图像。在放大因子为2时,在NWPU-RESISC45和UCMerced-LandUse两个测试集上得到的PSNR(峰值信噪比)和SSIM(结构相似性)分别为37.720 4 dB、37.999 6 dB和0.962 1、0.965 4,均优于DSSR、IRN和MPSR等先进的遥感图像超分辨率重建的模型,证明了所设计模型的有效性和鲁棒性。 |
| 关键词: 超分辨率重建;深度学习;多尺度特征;混合注意力机制 |
|
中图分类号: TP391
文献标识码: A
|
|
| Remote Sensing Image Reconstruction with Multi-scale Information and Hybrid Attention |
|
CAO Chunping, LI Ang
|
(School of Optical-Electrical & Computer Engineering, University of Shanghai f or Science & Technology, Shanghai 200093, China)
ccpgcd@163.com; leon_usst_365@163.com
|
| Abstract: This paper proposes a remote sensing image reconstruction model that integrates multi-scale information and hybrid attention, to address the problem of poor performance in reconstructing remote sensing images using a single size convolutional kernel. Two different multi-scale feature extraction blocks are used in the model, which can effectively extract high-frequency and low-frequency features in feature maps under different receptive fields. The weight of multi-scale features is adaptively adjusted through a hybrid attention mechanism, and the reconstruction module is used to reconstruct high-definition remote sensing images. When the amplification factor is 2, the PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) obtained on the NWPU-RESISC45 and UCMerced-LandUse test sets are 37.720 4 dB, 37.999 6 dB and 0.962 1, 0.965 4, respectively, which are superior to the advanced models for super-resolution reconstruction of remote sensing images such as DSSR, IRN, and MPSR, demonstrating the effectiveness and robustness of the proposed model. |
| Keywords: super-resolution reconstruction; deep learning; multi-scale feature; hybrid attention mechanism |
