WCSE

An Improved Capsule Network for the Detection of X-ray Security Images

Abstract— For X-ray security images inspection, the traditional convolutional neural networks require a large number of samples to train, but capsule network is not limited by training sample size, which contains the information of the position, size, and rotation angle of the object which is recognized through its unique capsule structure to achieve better training results with a small size data set. In this paper, we propose an improved capsule network named as DR-CapsNet for the X-ray security images inspection. DR-CapsNet uses DenseNet as the backbone network to extract image features more efficiently, and uses ResNet to connect the input and the output of DenseNet to prevent overfitting caused by deep network, then adds CBAM to the convolutional layer and the primary capsule layer, that can improve the learning ability, generalization ability and recognition accuracy. The experimental results show that the improved capsule network model DR-CapsNet has 99% recognition accuracy on the small size data set based on GDXray. Compared with the original capsule network, DR-CapsNet improves accuracy by 2%.

Index Terms— X-ray security inspection, Capsule network, DenseNet, ResNet, CBAM, small size data set.

Hong Zhang
School of Automation, Xi'an University of Posts and Telecommunications, China; Automatic Sorting Technology Research Center, State Post Bureau of the People’s Republic of China
Baoyang Liu
School of Automation, Xi'an University of Posts and Telecommunications, China; Automatic Sorting Technology Research Center, State Post Bureau of the People’s Republic of China

[Download]

Cite: Hong Zhang, Baoyang Liu, "An Improved Capsule Network for the Detection of X-ray Security Images," WCSE 2022 Spring Event: 2022 9th International Conference on Industrial Engineering and Applications, pp. 105-112, Sanya, China, April 15-18, 2022.