基于改进YOLOX算法的X射线图像违禁品检测方法  

Detection of prohibited items in X-ray images based on modified YOLOX algorithm

作　　者：袁金豪 张南峰 阮洁珊 高向东[1] YUAN Jinhao;ZHANG Nanfeng;RUAN Jieshan;GAO Xiangdong(Guangdong Provincial Welding Engineering Technology Research Center,Guangdong University of Technology,Guangzhou 510006,China;Huangpu Customs Technical Center,Dongguan 523076,China)

机构地区：[1]广东工业大学广东省焊接工程技术研究中心,广州510006 [2]黄埔海关技术中心,东莞523076

出　　处：《激光技术》2023年第4期547-552,共6页Laser Technology

基　　金：广州市技术创新发展专项基金资助项目(202002020068)。

摘　　要：为了实现自动检测X射线图像中的违禁品,解决相互遮挡、目标相近和小目标违禁品检测难的问题,提出一种基于改进的你只观察一次(YOLOX)算法的X射线图像违禁品检测方法。首先在YOLOX的主干网络低层中引入使用大核注意力构建的空间注意力,提取低层特征图的远距离依赖信息和纹理信息,之后在主干网络的中层和高层增加卷积块的注意力模块以增强感兴趣区域信息并抑制无用信息;该方法在公开的安全检查X射线数据集上进行实验,同时为改善模型的鲁棒性,在训练前70个周期使用Mosaic数据增强方法。结果表明,改进的模型较基本模型增加少量的参数和计算量,均值平均精度增加2.45%,提升到87.88%,平均推理速率为58.5 frame/s。该研究为即时自动检测X射线图像中违禁品提供了有益的参考。In order to realize automatic detection of contraband in X-ray images and to work out troubles of detecting mutual shaded,close and small-target prohibited items,an meliorative detection method on the strength of you only look once(YOLOX)algorithm was presented.Firstly,the spatial attention constructed with large kernel attention was introduced in the lower layer of YOLOX backbone network to extract the long-distance dependence information and texture message of the feature map in the lower layer.Then,the convolution block attention module was inserted in the middle and high layer of YOLOX backbone network to heighten the region of interest information and restrain unnecessary information.The proposed means was experimented on an overt security inspection X-ray dataset,meanwhile,in order to strengthen the robustness of the model,Mosaic data augmentation was used in the first 70 training epoch.The results show that,comparing with the basic model,the improved model put on a small amount of parameters and calculations.Mean average precision increases by 2.45%to 87.88%,and average inference velocity is 58.5 frames/s.This study can provide a salutary reference for automatically immediate detection of prohibited items in X-ray images.

关 键 词：X射线光学 违禁品检测 YOLOX算法 大核注意力 空间注意力 卷积块的注意力模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]