Title: PMix: a method to improve the classification of X-ray prohibited items based on probability mixing

Authors: Qitong Lu; Peng Han; Jian Qiu; Kunyuan Xu; Kaiqing Luo; Dongmei Liu; Li Peng

Addresses: Guangdong Provincial Engineering Research Centre for Optoelectronic Instrument, School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou, Guangdong, China ' Guangdong Provincial Engineering Research Centre for Optoelectronic Instrument, School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou, Guangdong, China ' Guangdong Provincial Engineering Research Centre for Optoelectronic Instrument, School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou, Guangdong, China ' School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou, Guangdong, China ' Guangdong Provincial Engineering Research Centre for Optoelectronic Instrument, School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou, Guangdong, China ' Guangdong Provincial Engineering Research Centre for Optoelectronic Instrument, School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou, Guangdong, China ' Guangdong Provincial Engineering Research Centre for Optoelectronic Instrument, School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou, Guangdong, China

Abstract: In recent years, much more attention has been paid to the field of automatic unmanned security checks. However, it's difficult to collect X-ray images about different types of dangerous items, and the X-ray imaging is characterised by a perspective view with a limited observation angle, which easily result in the X-ray image blurring, information loss and mistakes of prohibited items recognition. Therefore, there are few data sets about X-ray dangerous goods images and data imbalance exists at the same time. To solve this problem, this paper proposes a probabilistic image mixed data processing method called PMix. This method can increase the data volume of positive samples without destroying the unbalanced proportion of data sets and improve the accuracy of the model. Compared to the baselines, the experiments show that the proposed method can effectively promote the accuracy of model classification by more than 5% under the condition of unbalanced data.

Keywords: X-ray; mixup; Cutmix; data imbalance; multi-label-classification.

DOI: 10.1504/IJWMC.2022.123318

International Journal of Wireless and Mobile Computing, 2022 Vol.22 No.2, pp.179 - 187

Received: 27 Oct 2021
Accepted: 26 Feb 2022
Published online: 08 Jun 2022 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article