苗硕, 李新伟, 杨艺,等.基于改进胶囊网络的X射线图像违禁品检测[J].河南理工大学学报（自然科学版）,2023,42(3):129-136.

MIAO S, LI X W, YANG Y,et al.Detection of contraband in X-ray images based on improved Capsule network[J].Journal of Henan Polytechnic University(Natural Science) ,2023,42(3):129-136.

基于改进胶囊网络的X射线图像违禁品检测

苗硕1, 李新伟1, 杨艺1, 王科平1, 崔科飞2

1.河南理工大学 河南省煤矿装备智能检测与控制重点实验室，河南 焦作  454000;2.郑煤机液压电控有限公司，河南 郑州  450016

摘要:针对X射线图像违禁品检测中存在的漏检和误检问题，提出基于改进胶囊网络的模型（DMF and SE Capsule）用于X射线图像违禁品检测。该模型在传统胶囊网络的基础上增加了特征增强（dilated convolution multi-scale feature fusion，DMF）模块和特征筛选（squeeze-and-excitation block，SE）模块。首先使用特征增强模块提取图像特征，通过增加空洞卷积层，并且将所得的高低层语义特征进行拼接融合，从而得到丰富的特征信息；然后再用特征筛选模块，以挤压激励的方式将得到的特征进行筛选；最后再经过网络的胶囊层，从而完成对违禁品的检测。为了验证模型对复杂场景下X射线图像中违禁品的检测能力，在SIXray数据集上进行实验，模型的检测准确率达到79.254%，与原始的胶囊网络（71.350%）相比提升了7.904%，因此，改进模型的检测能力提升明显。

关键词:违禁品检测;胶囊网络;空洞卷积;多特征融合;特征筛选

doi:10.16186/j.cnki.1673-9787.2021080065

基金项目:国家重点研发计划项目（2018YFC0604502）；河南省科技攻关计划项目（192102210099，212102210390）

收稿日期:2021/08/18

修回日期:2021/11/24

出版日期:2023/05/25

Detection of contraband in X-ray images based on improved Capsule network

MIAO Shuo1, LI Xinwei1, YANG Yi1, WANG Keping1, CUI Kefei2

1.Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment，Henan Polytechnic University，Jiaozuo  454000，Henan，China;2.Zhengmeiji Hydraulic Electric Control Co.，Ltd.，Zhengzhou  450016，Henan，China

Abstract:Aiming at the problems of missing and false detection in X-ray image contraband detection，a model based on improved Capsule network（DSCapsule）was proposed.Feature enhancement block（DMF）and feature screening block（SE）were added in this model.Firstly，feature enhancement block was used to extract image features，and feature information was obtained by adding dilated convolution layers and splicing the semantic features of high and low layers.Secondly，the feature screening block was used to screen the obtained features by Squeezing and Excitation.Finally，the detection of contraband was completed through the Capsule layer of the network.In order to verify the detection ability of the model for contraband in complex X-ray images，the model was verified on SIXray dataset.The accuracy of the model reached 79.254%，which was 7.904% higher than the original Capsule network（71.350%）.Therefore，the detection ability of the improved model was significantly improved.

Key words:contraband detection;Capsule network;dilated convolution;multi-feature fusion;feature selection

基于改进胶囊网络的X射线图像违禁品检测_苗硕.pdf