基于轻量化生成对抗网络的遥感图像超分辨率重建

Super-resolution reconstruction of remote sensing images based on light weight generative adversarial network

针对ESRGAN模型复杂度高、特征提取与表示性能欠佳的问题,提出了一种基于轻量化生成对抗网络(Light weight Generative Adversarial Network, LwGAN)的遥感图像超分辨率重建算法。该算法以改进残差密集模块(Improved Residual Dense Block, IRDB)为基础块构建生成网络的高阶特征提取部分,提取了丰富的多样化特征,同时建立了特征的通道及长距离位置关系,在降低模型参数量的同时提升了模型的特征提取与表示性能。通过在UC MERCED和NWPU-RESISC45数据集上的实验结果表明,与ESRGAN相比,LwGAN获取了更大的峰值信噪比和结构相似度,显著提升了遥感图像的超分辨率重建性能,可视化结果表明重建图像恢复了更多的纹理细节信息,同时模型参数量仅为原始ESRGAN的约三分之一,大幅地提高了模型的运行效率,为后续遥感图像的分析处理奠定了基础。

Aiming at the problems of high complexity and poor feature extraction and presentation performance of ESRGAN model,a super-resolution reconstruction algorithm based on Light weight Generative Adversarial Network(LwGAN)is proposed.The Improved Residual Dense Block(IRDB)is used as the base block to construct the high order feature extraction part of the generated network,extract rich and diversified features,and establish the feature channel and long-distance location relationship.In addition to reducing the number of model parameters,the feature extraction and presentation performance of the model are improved.The experimental results on UC MERCED and NWPU-RESISC45 datasets show that compared with ESRGAN,LwGAN obtains larger peak signal-to-noise ratio and structural similarity,significantly improves the performance of super-resolution reconstruction of remote sensing ima-ges,and the visualization results show that the reconstructed images recover more texture detail information,while the number of model parameters is only about one-third of that of the original ESRGAN,which significantly improves the operation efficiency of the model and lays the foundation for subsequent analysis and processing of remote sensing images,