改进的LapSRN遥感图像超分辨重建

Improved LapSRN for remote sensing image super-resolution reconstruction

为了进一步提高遥感图像的超分辨效果,获得细节信息更加丰富的遥感图像,对深度拉普拉斯金字塔网络(Deep Laplacian pyramid super-resolution network,DLPSRN)进行了改进。遥感图像的特征提取是超分辨过程中重要的一部分,直接影响遥感图像超分辨重建的效果,本文改进了深度拉普拉斯金字塔网络特征的提取模块,应用非对称卷积块即多个并行的卷积核代替单一卷积核,可获得更好的图像特征提取效果。同时,加入了自适应特征集成模块,对提取的不同分辨率的图像特征进行特征交互。预测图像特征的最终权重,按照最终权重来进行遥感图像的超分辨重建,可获得更好的超分辨重建效果。实验结果表明,与原算法相比,改进的深度拉普拉斯金字塔网络重建图像,主观视觉效果和客观评价指标均有所提升,主观视觉上图像细节增多,客观评价指标峰值信噪比提高了0.61 dB,结构相似性提高了0.076。

In order to further improve the super-resolution effect of remote sensing images and obtain remote sensing images with more detailed information,the Deep Laplacian pyramid super-resolution network(DLPSRN)is improved.The feature extraction of remote sensing images is an important part of the super-resolution process,which directly affects the effect of super-resolution reconstruction of remote sensing images.The feature extraction module of the Deep Laplacian pyramid network is improved,using asymmetric convolutional blocks,namely multiple parallel convolution kernels replace single convolution kernels to obtain better image feature extraction effects.At the same time,the adaptive feature integration module is used to interact with the extracted image features of different resolutions.The final weight of image features is predicted.The super-resolution reconstruction of remote sensing images is performed,according to the final weight to obtain better super-resolution and rebuild the effect.The experimental results show that compared with the original algorithm,the improved deep Laplacian pyramid network reconstruction image improves subjective visual effects and objective evaluation indicators,subjective visual details of the image increase,and the objective evaluation indicator peak signal-to-noise ratio increases by 0.61 dB,the structural similarity is improved by 0.076.