面向單幅遙感圖像超分辨率的空間自適應(yīng)及 頻率融合網(wǎng)絡(luò)

打開(kāi)文本圖片集

關(guān)鍵詞：遙感圖像；超分辨率；混合尺度特征；空頻互補(bǔ)信息中圖分類號(hào)：TP391 文獻(xiàn)標(biāo)識(shí)碼：Adoi：10.37188/OPE.20253308.1238 CSTR：32169.14.OPE.20253308.1238

Abstract： Most of the existing methods of remote sensing image super-resolution are unable to fully explore the self-similarity information at hybrid scales and the correlation between cross-scale regions.Moreover，these methods ignore the ability of the frequency domain to perceive the high-frequency information of the images. To addressthis problem，a Spatial Adaptation and Frequency Fusion Network （SAF2Net）

was proposed. Firstly， SAF2Net introduced a hybrid-scale spatially-adaptive feature modulation，which adopted a feature pyramid-like approach to obtain discriminative features at diferent scales and enriched the expression ability of multi-scale features.Subsequently，a global multi-scale field selection block was designed to extract the correlation features of cross-scale regions. On this basis，a spatial adaptively selection block and a frequency separation selection block were introduced to fuse space-frequency complementary information to enhance local features，improving the model's ability to model the high-frequency content of images.Multiple sets of experiments are conducted on two remote sensing image datasets，which indicates that the quantitative evaluation metrics obtained by SAF2Net outperform those of other comparative methods. Taking the UCMerced dataset with 3 times super-resolution as an example，the proposed method improves PSNR and SSIMby O.11 dB and O.OO3 3，respectively，in compared with the next best method HAUNet. In terms of the subjective visual quality， SAFNet is able to recover more clear texture details.The experimental results demonstrate that the SAF2Net proposed is capable of mining the hybridscale global information from two different perspectives as wellas fusing the space-frequency complementary features efectively，which exhibits competitive performance in the remote sensing image super-resolution task.

Key words： remote sensing image；super resolution；hybrid-scale features； space-frequency complementaryinformation

1引言

隨著遙感應(yīng)用的日益發(fā)展，遙感圖像提供了豐富的空間細(xì)節(jié)和地物特征，在土地覆蓋分割1、分類[2、目標(biāo)精確識(shí)別[3以及變化監(jiān)測(cè)[4等諸多研究領(lǐng)域中，扮演著至關(guān)重要的角色。（剩余24375字）