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

发布时间：2018-08-20 18:12

【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。
[Abstract]:A method of target detection based on hierarchical and multi-scale convolution feature extraction is proposed for image detection of tank armored targets. Based on the VGG-16 network, the structure and parameters of the network are modified and fine-tuned based on the transfer learning, and the proposed region extraction network and the target detection subnetwork are combined to realize the accurate detection of the target. In order to enhance the detection ability of small and weak targets, the proposed region is extracted from several convolutional feature maps with different resolution layers to enhance the detection ability of small and weak targets, and the target detection subnetwork is used to detect small and weak targets. A convolutional feature map with higher resolution is used to extract the target, and an additional upper sampling layer is added to improve the resolution of the feature image. Through the combination of multi-scale training, difficult negative sample mining and other design and training methods, the proposed method has achieved excellent detection results on the constructed tank armor target data set. The accuracy and speed of target detection are better than the current mainstream detection methods.
【作者单位】：装甲兵工程学院控制工程系;
【基金】：总装备部院校科技创新工程项目(ZXY14060014)
【分类号】：E919;TP391.41

【相似文献】