结合空间语义注意力的二段式遥感图像修复网络

A two-stage remote sensing image inpainting network combined with spatial semantic attention

高分辨率遥感图像的缺失区域中地物种类复杂多样、空间异质性高,导致图像修复结果中存在纹理模糊和结构扭曲的问题,且在边界和复杂纹理区域尤为突出。因此提出一种结合空间语义注意力的二段式遥感图像修复网络。该网络由粗修复网络和精修复网络串联而成,旨在使用粗略修复网络提供的先验信息,引导精修复网络对缺失区域的复原。在粗修复网络中,构建多级损失结构以强化网络训练的稳定性;在精修复网络中,提出一种新的空间语义注意力机制,并依据网络特征的分布特点,区别性将空间语义注意力嵌入在编码器和解码器中,以确保局部特征的连续性和全局语义信息的相关性。实验结果表明,所提方法相比于现有其他算法可以进一步提升图像修复效果。

In high-resolution remote sensing images,missing areas feature intricate surface features and pronounced spatial heterogeneity,causing the image inpainting results to suffer texture blurring and structural distortion,particularly for boundaries and areas with complex textures.This study proposed a two-stage remote sensing image inpainting network combined with spatial semantic attention(SSA).The network comprised two networks in series:one for coarse image inpainting and one for fine-scale image inpainting(also referred to as the coarse and fine-scale networks,respectively).This network was designed to guide the fine-scale network to restore the missing areas using the priori information provided by the coarse network.In the coarse network,a multi-level loss structure was constructed to enhance the stability of network training.In the fine-scale network,a novel SSA mechanism was proposed,with SSA being embedded differentially in the encoder and decoder based on the distribution of network features.This ensured the continuity of local features and the correlation of global semantic information.The experimental results show that the network proposed in this study can further improve the image inpainting effects compared to other existing algorithms.