计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (1): 224-231.DOI: 10.3778/j.issn.1002-8331.2008-0345

• 图形图像处理 • 上一篇    下一篇

多通道区域建议的多尺度X光安检图像检测

康佳楠,张良   

  1. 中国民航大学 电子信息与自动化学院,天津 300300
  • 出版日期:2022-01-01 发布日期:2022-01-06

Multi-scale X-Ray Security Inspection Image Detection with Multi-channel Region Proposal

KANG Jianan, ZHANG Liang   

  1. College of Electronic Information and Automation, Civil Aviation University of China, Tianjin 300300, China
  • Online:2022-01-01 Published:2022-01-06

摘要: 针对安检X光图像检测中的违禁品尺度差异问题,对Faster RCNN网络进行改进,提出一种基于多通道区域建议网络(muiti-channel region proposal network,MCRPN)。考虑到不同层卷积特征在视觉语义上的互补性,进行多层特征提取,融合VGG16高层较丰富的语义特征和低层较浅的边缘特征;修改多通道RPN中的锚框参数,将生成的多尺度候选目标区域分别映射到对应的特征图上,构建多尺度违禁品检测网络;在多通道上引入膨胀卷积,设计一种多分支膨胀卷积模块(dilated convolutions module,DCM),增大感受野,增强不同尺度的特征。将改进的算法在自制数据集SIXray_OD上进行实验,检测的平均精度达到84.69%,测性能较原网络提高了6.28%。实验结果表明,改进算法的识别精度有一定提高。

关键词: 目标检测, Faster RCNN模型, 多尺度, 多通道, 膨胀卷积

Abstract: Aiming at the scale difference problem of prohibited items in X-ray images security inspection, the Faster RCNN network is improved and a multi-channel region proposal network(MCRPN) is designed on the basis of Faster RCNN. Firstly, considering the complementarity of different layers convolutional features in visual semantics, multi-layer feature extraction is introduced to combine richer high-level semantic features and low-level edge features of VGG16. Secondly, the anchor frame parameters in multi-channel RPN are modified. The generated multi-scale candidate regions are mapped to the corresponding feature maps, and a multi-scale contraband detection network is constructed. Finally, dilated convolution is introduced, and a dilated convolution module(DCM) is added to the multi channels to increase the receptive field, and the feature of different scales are enhanced. The improved algorithm is tested on the self-made data set SIXray_OD. The average detection accuracy reaches 84.69%, and the detection performance is improved by 6.28% compared with the original Faster RCNN network. The experimental results show that the recognition accuracy of the proposed algorithm has been improved.

Key words: object detection, Faster RCNN model, multi-scale, multi-channel, dilated convolution