基于U⁃net卷积神经网络的多尺度遥感图像分割算法  

Multi⁃scale remote sensing image segmentation algorithm based on U⁃net convolutional neural network

作　　者：刘丹英 刘晓燕[1] LIU Danying;LIU Xiaoyan(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500

出　　处：《现代电子技术》2023年第21期44-47,共4页Modern Electronics Technique

摘　　要：多尺度遥感图像的非本质特征量较大,不仅易导致图像噪声较大,也增加了图像分割的难度。为充分保留分割后多尺度遥感图像的边缘特征,在U⁃net卷积神经网络下提出新的图像分割算法。以U⁃net卷积神经网络为基网,提取被分割图像特征,获得被分割图像细节信息;计算相邻像素和原始像素特征向量的欧氏距离,结合去噪算法,通过归一化参数处理,建立相似性函数,实现对多尺度遥感图像分割特征增强处理;计算分割框候选偏差值;根据U⁃net卷积神经网络结构确定局部最优合并区域对;计算度量区域的距离,使用全局最优区域合并方法更新分割时间复杂度,实现多尺度遥感图像整体分割。由实验结果可知,该算法能够精准确定指定建筑物位置,并保留建筑物完整边缘细节信息。The large amount of non essential features in multiscale remote sensing images not only easily leads to large image noise,but also increases the difficulty of image segmentation.In order to fully preserve the edge features of segmented multiscale remote sensing images,a new image segmentation algorithm based on U⁃net convolutional neural network is proposed.The U⁃net convolutional neural network is used as the base network to extract the features of the segmented image and obtain the detailed information of the segmented image.The Euclidean distance between adjacent pixels and the original pixel feature vector is calculated.In combination with a denoising algorithm,a similarity function is established by normalization parameter processing to achieve segmentation feature enhancement processing for multiscale remote sensing images.The segmentation frame candidate deviation value is calculated.The local optimal combination region pair is determined according to the U⁃net convolutional neural network structure.The distance of the measurement region is calculated,and the segmentation time complexity is updated with the global optimal region merging method,so as to achieve the overall segmentation of multiscale remote sensing images.From the experimental results,it can be seen that the proposed algorithm can determine the location of a specified building accurately and preserve the full edge detail information of the building.

关 键 词：U⁃net卷积神经网络 特征提取 相邻像素 相似性函数 分割框候选偏差 多尺度 遥感图像 分割 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]