计算机科学 ›› 2021, Vol. 48 ›› Issue (1): 175-181.doi: 10.11896/jsjkx.200200023

• 计算机图形学与多媒体 • 上一篇    下一篇

基于域适应的X光图像的目标检测

何彦辉1, 吴桂兴1,2, 吴志强1   

  1. 1 中国科学技术大学软件学院 江苏 苏州 215123
    2 中国科学技术大学苏州研究院 江苏 苏州 215123
  • 收稿日期:2020-02-05 修回日期:2020-05-24 出版日期:2021-01-15 发布日期:2021-01-15
  • 通讯作者: 吴桂兴(gxwu@ustc.edu.cn)
  • 作者简介:sa517114@mail.ustc.edu.cn
  • 基金资助:
    江苏省自然科学基金(BK20141209)

Domain Alignment Based Object Detection of X-ray Images

HE Yan-hui1, WU Gui-xing1,2, WU Zhi-qiang1   

  1. 1 School of Software Engineering,University of Science and Technology of China,Suzhou,Jiangsu 215123,China
    2 Suzhou Research Institute,University of Science and Technology of China,Suzhou,Jiangsu 215123,China
  • Received:2020-02-05 Revised:2020-05-24 Online:2021-01-15 Published:2021-01-15
  • About author:HE Yan-hui,born in 1994,postgra-duate.His main research interests include object detection,transfer learning and so on.
    WU Gui-xing,born in 1972,Ph.D,professor,Ph.D supervisor,researcher,is a member of China Computer Federation.His main research interests include information theory,multimedia and so on.
  • Supported by:
    Natural Science Foundation ofJiangsu Province,China(BK20141209).

摘要: 随着卷积神经网络的发展,X光安全检查图像的自动目标检测算法已经取得了重大进步。但是,当将这些目标检测算法应用到不同于训练集数据的新数据,即训练域数据和测试域数据的图像数据服从不一致的分布时,这些检测算法的性能通常会降低。根据X光成像的变化,提出一种基于上下文的透射率自适应域对齐方法,用于解决检测算法的域不适应问题。首先,通过利用X光图像中存在的颜色信息,设计了一种注意力机制来分别处理X光图像的每个颜色通道特征,解决不同X光机器的颜色差异问题。接着,提出一种多分辨率特征对齐方法,以解决不同厂商不同X光图像之间的数据分布差异。最后,使用上下文向量作为对抗训练的正则化,利用邻域信息提高测试精度。基于X光图像数据集和Cityscape数据集的实验表明,所提方法解决了目标检测算法在不同于训练域数据的测试域中精度下降的问题。

关键词: 目标检测, X光图像, 域适应, 透射率适应, 卷积神经网络

Abstract: Significant progress has been made towards building accurate automatic object detection systems for a variety of parcel security check applications using convolutional neural networks.However,the performance of these systems often degrades when they are applied to new data that differs from the training data,for example,due to variations in X-ray imaging.In this paper,we propose a context-based and transmittance adaptive domain alignment method to address the above performance degradation.Firstly,by using color information existed in X-ray images,we design an attention mechanism to process each color channel of an X-ray image separately to solve the problem of color differences among different X-ray machines.Next,we develop a feature alignment method to reduce the statistics difference among different X-ray images generated by various manufacturers.Finally,we propose to use a context vector as a regularization for the improvement of adversarial training to improve the precision.Theme-thod proposed in this paper solves the problem of the accuracy degradation of the object detection in the test domain,which is different from the training domain.

Key words: Object detection, X-ray images, Domain alignment, Transmittance-adptation, Convolutional neural network

中图分类号: 

  • TP183
[1] EVERINGHAM M,VAN GOOL L,WILLIAMS C K,et al.The pascal visual object classes(voc) challenge [J].IJCV 2010,88(2):303-338.
[2] REN S,HE K,GIRSHICK R,et al.Faster rcnn:Towards real-time object detection with region proposal networks[C]//NIPS.2015:1-9.
[3] MERY D,RIFFO V,ZSCHERPEL U,et al.Gdxray:The database of x-ray images for nondestructive testing[J].Journal of Nondestructive Evaluation,2010,34(4):42.
[4] MIAO C J,XIE L X,WAN F,et al.SIXray:A Large-scale Security Inspection X-ray Benchmark for Prohibited Item Discovery in Overlapping Images[J].arXiv:1901.00303.
[5] Ailibaba.Dataset :Logistics goods restriction monitoring[EB/OL].(2019-03-05) [2019-06-08].https://tianchi.aliyun.com/competition/entrance/231703/introduction.
[6] GANIN Y,LEMPITSKY V.Unsupervised domain adaptationby backpropagation[C]//ICML.2014.
[7] BEN-DAVID S,BLITZER J,CRAMMER K,et al.A theory of learning from different domains[J].Machine Learning,2010,79(1):151-175.
[8] TZENG E,HOFFMAN J,SAENKO K,et al.Adversarial dis-criminative domain adaptation[C]//CVPR.2017:11,19.
[9] TZENG E,HOFFMAN J,ZHANG N,et al.Deep domain confusion:Maximizing for domain invariance[J].arXiv:1412.3474.
[10] LONG M,CAO Y,WANG J,et al.Learning transferable fea-tures with deep adaptation networks[C]//ICML.2015.
[11] DUAN L,TSANG I W,XU D.Domain transfer multiple kernel learning[J].TPAMI,2012,34(3):465-479.
[12] INOUE N,FURUTA R,YAMASAKI T,et al.Cross-domainweakly-supervised object detection through progressive domain adaptation[J].arXiv:1803.11365.
[13] ZHU J Y,PARK T,ISOLA P,et al.Unpaired image-to-image translation using cycle-consistent adversarial networks[J].arXiv:1703.10593.
[14] CHEN Y,LI W,SAKARIDIS C,et al.Domain adaptive faster r-cnn for object detection in the wild[J].arXiv:1803.03243.
[15] SAITO K,USHIKU Y,HARADA T,et al.Strong-Weak Distribution Alignment for Adaptive Object Detection[C]//CVPR.2019:2154-2162.
[16] ZHU X G,JIANG M,YANG C Y.Adapting Object Detectors via Selective Cross-Domain Alignment [C]//CVPR.2019:687-696.
[17] SCHWAB P,KELLER E,MUROI C.Not to Cry Wolf:Distantly Supervised Multitask Learning in Critical Care[J].arXiv:1802.05027.
[18] CAI Z,VASCONCELOS N.Cascade R-CNN:Delving into high quality object detection[J].arXiv:1712.00726.
[19] HE K M,ZHANG X Y,REN S Q,et al.Deep Residual Learning for Image Recognition[J].arXiv:1512.03385.
[20] LIN T Y,GOYAL P,GIRSHICK R,et al.Focal loss for dense object detection[J].arXiv:1708.02002.
[21] LIN T Y,DOLLAR P,GIRSHICK R,et al.Feature pyramidnetworks for object detection[C]//CVPR.2017:2117-2125.
[22] DENG J,DONG W,SOCHER R,et al.Imagenet:A large-scale hierarchical image database[C]//CVPR.2009:248-255.
[23] CORDTS M,OMRAN M,RAMOS S,et al.The Cityscapesdataset for semantic urban scene understanding[C]//CVPR.2016:268-276.
[1] 单美静, 秦龙飞, 张会兵. L-YOLO:适用于车载边缘计算的实时交通标识检测模型[J]. 计算机科学, 2021, 48(1): 89-95.
[2] 李亚男, 胡宇佳, 甘伟, 朱敏. 基于深度学习的miRNA靶位点预测研究综述[J]. 计算机科学, 2021, 48(1): 209-216.
[3] 张佳嘉, 张小洪. 多分支卷积神经网络肺结节分类方法及其可解释性[J]. 计算机科学, 2020, 47(9): 129-134.
[4] 朱玲莹, 桑庆兵, 顾婷婷. 基于视差信息的无参考立体图像质量评价[J]. 计算机科学, 2020, 47(9): 150-156.
[5] 崔彤彤, 王桂玲, 高晶. 基于1DCNN-LSTM的船舶轨迹分类方法[J]. 计算机科学, 2020, 47(9): 175-184.
[6] 齐少华, 徐和根, 万友文, 付豪. 动态环境下的语义地图构建[J]. 计算机科学, 2020, 47(9): 198-203.
[7] 刘海潮, 王莉. 基于深度图卷积胶囊网络的图分类模型[J]. 计算机科学, 2020, 47(9): 219-225.
[8] 袁晨晖, 程春玲. 基于PE散度实例过滤的深度域适应方法[J]. 计算机科学, 2020, 47(8): 151-156.
[9] 姚兰, 赵永恒, 施雨晴, 于明鹤. 一种基于视频分析的高速公路交通异常事件检测算法[J]. 计算机科学, 2020, 47(8): 208-212.
[10] 梁正友, 何景琳, 孙宇. 一种用于微表情自动识别的三维卷积神经网络进化方法[J]. 计算机科学, 2020, 47(8): 227-232.
[11] 孙彦丽, 叶炯耀. 基于剪枝与量化的卷积神经网络压缩方法[J]. 计算机科学, 2020, 47(8): 261-266.
[12] 程哲, 白茜, 张浩, 王世普, 梁宇. 使用深层卷积神经网络提高Hi-C 数据分辨率[J]. 计算机科学, 2020, 47(6A): 70-74.
[13] 赫磊, 邵展鹏, 张剑华, 周小龙. 基于深度学习的行为识别算法综述[J]. 计算机科学, 2020, 47(6A): 139-147.
[14] 孙正, 王新宇. 深度学习在光声成像中的应用现状[J]. 计算机科学, 2020, 47(6A): 148-152.
[15] 张曼, 李杰, 丁荣莉, 成昊天, 沈霁. 基于改进YOLO-V2算法的遥感图像目标检测技术研究[J]. 计算机科学, 2020, 47(6A): 176-180.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
[1] 雷丽晖,王静. 可能性测度下的LTL模型检测并行化研究[J]. 计算机科学, 2018, 45(4): 71 -75 .
[2] 文俊浩,孙光辉,李顺. 基于用户聚类和移动上下文的矩阵分解推荐算法研究[J]. 计算机科学, 2018, 45(4): 215 -219 .
[3] 张昱, 高克宁, 于戈. 一种融合节点属性信息的社会网络链接预测方法[J]. 计算机科学, 2018, 45(6): 41 -45 .
[4] 池凯凯, 林一民, 李燕君, 程珍. 能量捕获传感网中吞吐量最大化的占空比方案[J]. 计算机科学, 2018, 45(6): 100 -104 .
[5] 陈嵘, 李鹏, 黄勇. 基于多特征融合的运动阴影去除算法[J]. 计算机科学, 2018, 45(6): 291 -295 .
[6] 孙海峰,宋丽丽. 路口中继辅助车载自组织网络路由算法[J]. 计算机科学, 2018, 45(5): 75 -78 .
[7] 许文浩,沈航,白光伟. 无线链路质量突变性感知的机会路由协议[J]. 计算机科学, 2018, 45(5): 83 -88 .
[8] 史经启,杨庚,孙彦珺,白双杰,闵兆娥. 支持浮点运算的高效并行全同态加密算法[J]. 计算机科学, 2018, 45(5): 116 -122 .
[9] 张雷,康宝生. 基于结构稀疏度和块差异度的目标移除图像修复[J]. 计算机科学, 2018, 45(5): 255 -259 .
[10] 刘枭, 王晓国. 基于概率图的银行电信诈骗检测方法[J]. 计算机科学, 2018, 45(7): 122 -128 .