基于深度学习的遥感图像地物分割方法  

Remote sensing image feature segmentation method based on deep learning

作　　者：沈言善 王阿川[1] SHEN Yan-shan;WANG A-chuan(College of Information and Computer Engineering, Northeast Forestry University, Harbin 150000, China)

机构地区：[1]东北林业大学信息与计算机工程学院,黑龙江哈尔滨150000

出　　处：《液晶与显示》2021年第5期733-740,共8页Chinese Journal of Liquid Crystals and Displays

基　　金：黑龙江省自然科学基金(No.C201414);哈尔滨市优秀学科带头人基金(No.2014RFXXJ040)。

摘　　要：针对传统方法难以同时有效地进行多目标分割,现有基于全卷积神经网络的地物分类方法在复杂场景下分类精度不高的问题,提出了一种基于U-net改进的全卷积神经网络DL-Unet,实现了对遥感图像不同类别地物的有效分割。该网络改进了传统的卷积方式,引入扩张卷积,在不增加网络参数的同时增大感受野。针对遥感影像中地物类别不均衡的问题,使用带权重的交叉熵作为模型的损失函数,有效克服了模型的选择偏好。对预测结果采用了相对多数投票策略进一步提高了各个地物类别的像素准确率(PA)。实验结果表明,该模型在各地物类别的平均像素准确率(MPA)和均交并比(MIoU)上相较于经典U-net网络分别提高了5.94%和9.45%,实现了对地物信息的有效分割。In view of the difficulty of traditional methods for simultaneous and effective multi-target segmentation,the existing ground object classification methods based on fully convolutional neural networks have low classification accuracy in complex scenes,this paper proposes an improved encoder-decoder based on the U-shaped network structure DL-Unet,which realizes the effective segmentation of remote sensing images.This network improves the traditional convolution method and introduces the expanded convolution,which increases the receptive field without increasing the network parameters.Aiming at the problem of imbalance in the clssification of featares in remote sensing images,weighted cross-entropy is used as the loss function of the model,which effectively overcomes the selection preference of the model function of the model effectively.The relative majority voting strategy is adopted for the prediction results to further improve the pixel accuracy(PA)of each feature category.The experimental results show that the mean pixel accuracy(MPA)and mean intersection over union(MIoU)of this model are improved by 5.94%and 9.45%respectively compared with the classic U-net,which verifies that the method in this paper is an effective remote sensing image classification method.

关 键 词：深度学习 扩张卷积 加权交叉熵 集成学习 语义分割 图像处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]