基于多层级特征融合的线状柔性体分割方法  

Deformable linear objects segmentation method based on multi-level feature fusion

作　　者：张长勇 李玉洲 张轩铖 ZHANG Chang-yong;LI Yu-zhou;ZHANG Xuan-cheng(College of Electronic Information and Automation,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学电子信息与自动化学院,天津300300

出　　处：《计算机工程与设计》2024年第10期3103-3110,共8页Computer Engineering and Design

基　　金：中国交通教育研究会研究课题基金项目(JT2022YB207);教育部产学合作协同育人基金项目(220700782044600);实验技术创新基金项目(2023CXJJ57)。

摘　　要：为解决线状柔性体分割精度低、速度慢的问题,提出一种改进DeepLabV3+网络。利用轻量化且易收敛的MobileNetV2作为主干特征提取网络,通过CA注意力模块对输入特征进行关键信息的集中关注,提出一种多尺度空洞金字塔池化结构,提升网络的感受野和模型训练效率,在解码层提出改进的级联特征融合模块融合三层浅层特征,提升语义信息的表征能力。实验结果表明,改进网络相比未改进网络MIOU、IOU分别提高2.82%、3.46%,预测时间减少5.2 ms,能够实现复杂背景下线状柔性体的准确分割。To solve the problem of low accuracy and low speed of linear flexible body segmentation,an improved DeepLabV3+network was proposed.The MobileNetV2,a lightweight and easy-to-converge feature extraction network,was utilized as the backbone network.The CA(channel attention)module was employed to concentrate on key information in the input features.A multi-scale atrous pyramid pooling structure was introduced to enhance the network’s receptive field and training efficiency.In the decoding stage,an improved cascaded feature fusion module was proposed to fuse three shallow-level features,thereby improving the representation capability of semantic information.Experimental results show that compared with the non-improved network,the improved network improves MIOU and IOU by 2.82%and 3.46%,respectively,and reduces the prediction time by 5.2 ms,which can realize the accurate segmentation of deformable linear objects under complex background.

关 键 词：语义分割 轻量化网络 注意力机制 特征融合 线状柔性体分割 空洞卷积 级联特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]