摘要
目的:针对多战场环境中军事人员目标较小以及军事人员与环境背景难以区分导致分割效果较差的问题,提出一种分割神经网络。方法:以U-Net为主干神经网络,基于空洞空间金字塔池化模块和双特征交叉融合模块提出编码-解码神经网络ASPP-DFCF-U-Net(以下简称“AD-U-Net”)。将多环境迷彩分割数据集(Multi Environment Camouflage Dataset,MECD)按比例分为训练集、验证集和测试集并进行数据增强,在MECD上对AD-U-Net进行训练并测试。为验证AD-U-Net的有效性,采用平均交并比、召回率、精确度、F1分数指标将AD-U-Net与U-Net、SegNet、FCN-8s 3种神经网络的分割结果进行对比分析。结果:AD-U-Net的平均交并比、召回率、精确度、F1分数分别为83.04%、89.58%、91.49%和90.52%,均高于U-Net、SegNet和FCN-8s,在目标较小、分割目标与背景高度相似的情况下具有更好的分割效果。结论:AD-U-Net在MECD上具有优良的分割效果,将其应用于搜救装备中进行人员搜救可大大提高军事人员的分割准确率,提高搜救效率。
Objective To propose a segmentation neural network to solve the problems of small military personnel targets in multiple battlefield environments and poor segmentation due to the difficulty of distinguishing military personnel from the environmental background.Methods An encoding-decoding neural network named ASPP-DFCF-U-Net(abbreviated as AD-U-Net)was proposed with U-Net as the backbone neural network,which was developed based on atrous spatial pyramid pooling module and dual feature cross fusion module.The Multi Environment Camouflage Dataset(MECD)was scaled into training,validation and test sets and augmented with data,and AD-U-Net was trained and tested on MECD.AD-U-Net had its effectiveness verified by comparing its segmentation results with those by U-Net,SegNet and FCN-8s,using the indexes of mean intersection over union,recall rate,precision and F1 score.Results The mean intersection over union,recall rate,precision and F1 score of AD-U-Net were 83.04%,89.58%,91.49%and 90.52%respectively,all higher than those of U-Net,SegNet and FCN-8s.It's proved AD-U-Net behaved well in case of smalltargets and highly similar segmentation targets and backgrounds.Conclusion AD-U-Net gains excellent segmentation effects on MECD and its application to the search and rescue equipment can greatly improve the segmentation accuracy of military personnel image and increase the efficiency of search and rescue.
作者
陶志文
张伟
周旗开
牛福
TAO Zhi-wen;ZHANG Wei;ZHOU Qi-kai;NIU Fu(Academy of Systems Engineering of Academy of Military Science of Chinese PLA,Beijing 100166,China)
出处
《医疗卫生装备》
CAS
2021年第10期7-11,共5页
Chinese Medical Equipment Journal
关键词
战场环境
军事人员图像分割
编码-解码神经网络
空洞空间金字塔池化
双特征交叉融合
battlefield environment
military personnel image segmentation
encoder-decoder neural network
atrous spatial pyramid pooling
dual feature cross fusion