一种基于分层多尺度卷积特征提取的坦克装甲目标图像检测方法  

Image Detection Method for Tank and Armored Targets Based on Hierarchical Multi-scale Convolution Feature Extraction

作　　者：孙皓泽 常天庆[1] 王全东[1] 孔德鹏[1] 戴文君[1] 

机构地区：[1]装甲兵工程学院控制工程系,北京100072

出　　处：《兵工学报》2017年第9期1681-1691,共11页Acta Armamentarii

基　　金：总装备部院校科技创新工程项目(ZXY14060014)

摘　　要：针对坦克装甲目标的图像检测任务,提出一种基于分层多尺度卷积特征提取的目标检测方法。采用迁移学习的设计思路,在VGG-16网络的基础上针对目标检测任务对网络的结构和参数进行修改和微调,结合建议区域提取网络和目标检测子网络来实现对目标的精确检测。对于建议区域提取网络,在多个不同分辨率的卷积特征图上分层提取多种尺度的建议区域,增强对弱小目标的检测能力;对于目标检测子网络,选用分辨率更高的卷积特征图来提取目标,并额外增加了一个上采样层来提升特征图的分辨率。通过结合多尺度训练、困难负样本挖掘等多种设计和训练方法,所提出的方法在构建的坦克装甲目标数据集上取得了优异的检测效果,目标检测的精度和速度均优于目前主流的检测方法。A target detection method based on hierarchical multi-scale convolution feature extraction is proposed for the image detection of tank and armored targets. The idea of transfer learning is used to modify and fine-tune the structure and parameters of VGG-16 network according to the target detection task,and the region proposal network and the detection sub-network are combined to realize the accurate detection of targets. For the region proposal network,the multi-scale proposals are extracted from the convolution feature maps of different resolutions to enhance the detection capability of small targets. For the object detection sub-network,the feature maps with high-resolution convolution are used to extract the targets,and an upsampling layer is added to enhance the resolution of the feature maps. With the help of multi-scale training and hard negative sample mining,the proposed method achieves the excellent results in the tank and armored target data set,and its detection accuracy and speed are better than the those ofcurrent mainstream detection methods.

关 键 词：兵器科学与技术 目标探测与识别 卷积神经网络 坦克装甲目标 目标检测 

分 类 号：TP391.413[自动化与计算机技术—计算机应用技术]