摘要
针对复杂环境下远距离目标在语义分割时易出现的边界模糊、断裂及目标丢失等问题,基于DeepLabV3+网络提出了一种结合边界信息的语义分割模型。该模型采用改进的Darknet-53网络代替原DeepLabV3+特征提取网络以加快模型运行速度,并设计了一种特征融合模块作为低层特征用于解码阶段恢复细节信息,为了进一步优化目标边界,利用特征共享原则,设计一种通过主体网络特征共享层学习多尺度信息以预测目标边界的边界提取模块,以此对分割图像进行约束优化,提升模型在边界处的预测准确率。实验结果表明,提出的语义分割模型能够有效缓解远距离目标语义分割时的边界模糊等问题。
To solve the problems of fuzzy boundary,fracture and target loss in semantic segmentation of long-distance targets in complex environment,a semantic segmentation model using boundary information based on DeepLabV3+ network is proposed.The improved Darknet-53 network is used to replace the original DeepLabV3+ feature extraction network to speed up the model’s operation,and a feature fusion module is designed as a low-level feature to recover the detailed information in the decoding stage.In order to further optimize the target’s boundary,by using the principle of feature sharing,a boundary extraction module is designed to predict the target’s boundary by learning multi-scale information through the feature sharing layer of the main network,so as to optimize the segmented image and improve the prediction accuracy of the model at the boundary.The experimental results show that the proposed semantic segmentation model can effectively alleviate the problems of fuzzy boundary,fracture and target loss in the semantic segmentation of long-distance targets.
作者
喻根
崔炜
徐照翔
刘馨柔
YU Gen;CUI Wei;XU Zhaoxiang;LIU Xinrou(Changchun University of Science and Technology,Changchun 130033,China)
出处
《电光与控制》
CSCD
北大核心
2021年第1期66-70,共5页
Electronics Optics & Control
基金
中国吉林省科学技术计划发展项目(20180201042G X)
2019年省预算内基本建设资金(2019C054-b)。
