遥感图像中细弱目标分割方法  

Segmentation method of thin and weak targets in remote sensing images

作　　者：叶秀芬[1] 于淼 郭书祥[1] 梁洪[1] YE Xiufen;YU Miao;GUO Shuxiang;LIANG Hong(School of Automation Harbin Engineering University, Harbin 150001, China)

机构地区：[1]哈尔滨工程大学自动化学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2020年第11期1689-1694,共6页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(41876100,61633004);科技部重点研发计划项目(2018YFC0310102,2017YFC0306002);中央高校基本科研业务费专项资金项目(HEUCFP201707).

摘　　要：为了解决现有图像分割方法在遥感图像细弱目标分割任务中所存在的分割精度较低,受背景噪声干扰明显等问题,本文提出了一种基于Dense-Unet网络和集成学习的改进语义分割方法,实现了高分辨率遥感图像中的高压电线一类细弱目标的精准分割。融合DenseNet思想和U-net网络结构,提出了一种Dense-Unet网络模型,从而提取图像中更加精细的特征;为了增强模型对小类别目标的学习能力,使用了代价敏感权重向量方法,改进了训练时的目标函数,提升了模型训练时的收敛速度和在测试数据上的泛化能力;采用了集成学习Bagging的方法对多个网络模型进行集成,采用决策级融合方法集成多个模型对预测结果进行投票表决,使得细弱目标分割精度得到进一步提升。在最终的测试集上进行网络模型评估表明:本文所提出算法的精度达到94%,在细弱目标分割效果上优于目前常用的深度学习语义分割网络,且拥有更好的泛化能力和优异的视觉效果。The existing image segmentation methods still exhibit several problems such as low segmentation accuracy and an obvious interference from background noise in the task of the segmentation of thin and weak targets in a remote sensing image.In response,this study proposes an improved semantic segmentation method based on Dense-U-net and ensemble learning to achieve the accurate segmentation of thin and weak targets in high-resolution remote sensing images.First,a DenseNet idea and a U-net network structure is combined to establish a new Dense-U-net model to extract detailed image features.Second,to enhance the learning ability of the model for small categories of targets,the cost-sensitive weight-vector method is used to improve the objective function during training.This consequently enhances convergence speed during training and the generalization ability of the model on test data.Finally,the ensemble learning method,Bagging,is adopted to integrate multiple network models by the decision-level fusion method to vote the predicted results.This further improves the segmentation accuracy of thin and weak targets.The network model evaluation of the final test set shows that the accuracy of the proposed algorithm in this study reaches 94%,which is superior to the deep-learning semantic segmentation network commonly used currently in terms of the segmentation effect of weak targets.Furthermore,the algorithm presented in this study exhibits better generalization ability and excellent visual effect.

关 键 词：遥感图像 深度学习 图像分割 语义分割 细弱目标分割 图像类别信息非均衡 集成学习 多模型融合 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]