顾及边缘结构优化的深度卷积神经网络超分辨率重建  

Edge structure optimization in deep convolutional neural network for super-resolution reconstruction

作　　者：邓盛满 DENG Shengman(Fujian Surveying and Mapping Institute,Fuzhou,Fujian 350003,China)

机构地区：[1]福建省测绘院,福建福州350003

出　　处：《北京测绘》2025年第4期561-566,共6页Beijing Surveying and Mapping

基　　金：福建省自然资源科技创新项目(KY-100000-04-2023-006)。

摘　　要：针对深度学习网络模型下的超分辨率重建(SRCNN)影像存在边缘结构不够清晰与纹理细节缺乏的问题,本文提出了一种顾及边缘结构优化的深度卷积神经网络(CNN)超分辨率重建方法。首先,通过引导滤波将原始遥感影像分离为边缘结构与平滑信息;其次,将分离的边缘结构与平滑信息输入到主干卷积神经网络模型,学习高、低分辨率特征向量间的非线性映射参数;最后,结合信息熵与均方误差作为卷积神经网络的损失函数,优化重建后的高分辨率图像纹理细节。实验结果表明,本文提出的方法能够改善重建影像的边缘结构,提升重建影像的纹理细节,从而实现更高质量的影像重建效果。This paper proposed a deep convolutional neural network(CNN) method for super-resolution reconstruction(SRCNN) that addressed the issues of unclear edge structures and lack of texture details in images reconstructed by deep learning models.First,the original remote sensing image was separated into edge structures and smooth information using guided filtering.Next,the separated edge structures and smooth information were input into the backbone CNN model to learn the nonlinear mapping parameters between high-and low-resolution feature vectors.Finally,the loss function of the CNN was optimized by combining information entropy and mean square error to improve the texture details in the reconstructed high-resolution image.Experimental results show that the proposed method can enhance the edge structure of the reconstructed image and improve its texture details,thus achieving higher-quality image reconstruction.

关 键 词：超分辨率重建(SRCNN) 卷积神经网络(CNN) 边缘结构优化 损失函数 

分 类 号：P258[天文地球—测绘科学与技术]