Journal of Computer Applications ›› 2018, Vol. 38 ›› Issue (12): 3414-3418.DOI: 10.11772/j.issn.1001-9081.2018050989

Previous Articles     Next Articles

X-ray security inspection method using active vision based on Q-learning algorithm

DING Jingwen, CHEN Shuyue, LU Guirong   

  1. School of Information Science & Engineering, Changzhou University, Changzhou Jiangsu 213164, China
  • Received:2018-05-11 Revised:2018-07-02 Online:2018-12-10 Published:2018-12-15
  • Contact: 陈树越
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (51176016), the Changzhou Engineering Technology Research Center Project (CM20179060).

基于Q学习算法的X光主动视觉安检方法

丁静文, 陈树越, 陆贵荣   

  1. 常州大学 信息科学与工程学院, 江苏 常州 213164
  • 通讯作者: 陈树越
  • 作者简介:丁静文(1993-),女,江苏如皋人,硕士研究生,主要研究方向:数字图像处理;陈树越(1963-),男,河北定州人,教授,博士,主要研究方向:机器视觉、图像处理、信号检测;陆贵荣(1968-),男,甘肃定西人,副教授,博士,主要研究方向:过程控制中传感器检测、电子仪器设计。
  • 基金资助:
    国家自然科学基金资助项目(51176016);常州市工程技术研究中心项目(CM20179060)。

Abstract: In order to solve the problems of poor detection performance and slow detection speed of the active vision security inspection method, a Heuristically Accelerated State Backtracking Q-Learning (HASB-QL) algorithm based on Q-Learning (QL) algorithm was proposed to estimate the next-best-view. The cost function and heuristic function were introduced by the proposed algorithm to improve the learning efficiency and speed up the convergence of QL. Firstly, the single view detection of X-ray images obtained by security scanner was performed. Secondly, the pose was estimated and the best rotation angle was obtained by comparing the selection strategy of repeated action in the state backtracking process, and then the single view detection was performed again until the threat object was detected. Moreover, the geometric constraint was established to eliminate the false alarms when the number of views was more than one in detection process. The X-ray images of handguns and razor blades in GDXray data set were used for the experiment. The experimental results show that, compared with active vision algorithm based on QL, the weighted average value of FF1 between the precision and recall of detecting the handguns by the improved active vision algorithm is increased by 9.60% and the detection speed is increased by 12.45%, while the F1 of detecting razor blades is increased by 2.51% and the detection speed is increased by 17.39%. The proposed algorithm can improve the performance and speed of threat object detection.

Key words: X-ray image, security inspection, active vision, state backtracking, heuristic function, Q-Learning (QL)

摘要: 针对主动视觉安检方法检测性能不高和检测速度慢的问题,基于Q学习(QL)算法提出了采用状态回溯的启发式Q学习(HASB-QL)算法进行最佳视角估计。该算法引入代价函数和启发函数,提高了学习效率,加快了Q学习收敛。首先,对通过安检扫描仪获取的X光图像进行单视角检测;然后,对姿势作出估计并通过在状态回溯过程中比较重复动作的选择策略获取最佳旋转角度,再次进行单视角检测,直到检测到危险品;此外,在检测过程中多于一个视角时,建立几何约束以消除误报。对GDXray数据集中的手枪和剃刀刀片的X光图像进行实验,实验结果表明,相比于以Q学习为基础的主动视觉算法,改进的主动视觉算法检测手枪所得精确率和召回率之间的加权平均值F1值提高了9.60%,检测速度提高了12.45%;检测剃刀刀片所得的F1值提高了2.51%,速度提高了17.39%。所提算法提高了危险品检测的性能和速度。

关键词: X光图像, 安检, 主动视觉, 状态回溯, 启发函数, Q学习

CLC Number: