Welcome to Acta Armamentarii ! Today is Share:

Acta Armamentarii ›› 2017, Vol. 38 ›› Issue (9): 1681-1691.doi: 10.3969/j.issn.1000-1093.2017.09.003

• Paper • Previous Articles     Next Articles

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

SUN Hao-ze, CHANG Tian-qing, WANG Quan-dong, KONG De-peng, DAI Wen-jun   

  1. (Department of Control Engineering, Academy of Armored Force Engineering, Beijing 100072, China)
  • Received:2016-11-14 Revised:2016-11-14 Online:2017-11-03

Abstract: 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 mo-dify 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 of current mainstream detection methods. Key

Key words: ordnancescienceandtechnology, targetdetectionandidentification, convolutionalneuralnetwork, tankandarmoredtarget, targetdetection

CLC Number: