Abstract:

To meet the demand of high spatio-temporal resolution remote sensing images for power facilities safety monitoring and emergency management， a deep convolutional network-based remote sensing image fusion enhancement model for power facilities was proposed. Firstly， a deep convolutional network was designed， including encoder， Residual Attention （RA） mechanism block， substitution attention mechanism block and decoder. Secondly， the two-layer convolution and the residual block of fusion channel attention mechanism were improved to increase the network's attention to details and key features of images， and enhance the feature extraction capability of the network. Thirdly， the multi-channel substitution attention block was improved to make the network paying more attention to the details of images. As the result， the performance of high-resolution image fusion reconstruction was improved. Finally， the loss function composition of the model was improved， and the composite loss function consisting of content loss and visual loss was adopted to improve training effect of the model. Experimental results indicate that the proposed model has the performance of image fusion reconstruction better than other fusion models significantly， and the detail textures of predicted image closer to those of the real image. Compared with Multi-stage Feature Compensation NET （MFCNET） model， the proposed model has the Correlation Coefficient （CC） improved by 1.6%. and the SSIM （Structure Similarity Index Measure） improved by 18.4%. It can be seen that the proposed model provides a basis for remote sensing image processing， especially for high-resolution reconstruction of small target remote sensing images.

Key words: remote sensing image, deep convolutional network, fusion enhancement, power facility, attention mechanism

摘要：

针对电力设施安全监测及应急管理对于高时空分辨率遥感图像需求，提出一种基于深度卷积网络的电力设施遥感图像融合增强模型。首先，设计包含编码器、残差注意力机制（RA）模块、置换注意力机制模块和解码器的深度卷积网络；其次，改进双层卷积与融合通道注意力机制残差模块，以提高网络对于图像细节及关键特征的关注度，并增强网络的特征提取能力；再次，改进多通道置换注意力模块，使得网络能够更加关注图像细节，从而提升高分辨图像融合重建的性能；最后，改进深度学习网络的损失函数组成，采用由内容损失以及视觉损失组成的复合损失函数，从而提高模型的训练效果。实验结果表明，所提模型的图像融合重建效果明显优于其他融合模型，预测图像在细节纹理上更接近真实图像，与多级特征补偿网络（MFCNET）模型相比，所提模型的重建图像的相关系数（CC）提升了1.6%，结构相似性指数（SSIM）提升了18.4%。可见，所提模型为遥感图像处理，特别是小目标遥感图像高分辨重建提供了基础。

关键词: 遥感图像, 深度卷积网络, 融合增强, 电力设施, 注意力机制