摘要
针对主动视觉安检方法准确率低、速度慢,不适用于实时交通安检的问题,提出了八度卷积(OctConv)和注意力机制双向门控循环单元(GRU)神经网络相结合的X光安检图像分类方法。首先,利用八度卷积代替传统卷积,对输入的特征向量进行高低分频,并降低低频特征的分辨率,在有效提取X光安检图像特征的同时,减少了空间冗余。其次,通过注意力机制双向GRU,动态学习调整特征权重,提高危险品分类准确率。最后,在通用SIXRay数据集上的实验表明,对8000幅测试样本的整体分类准确率(ACC)、特征曲线下方面积(AUC)、正类分类准确率(PRE)分别为98.73%、91.39%、85.44%,检测时间为36.80 s。相对于目前主流模型,本文方法有效提高了X光安检图像危险品分类的准确率和速度。
Due to the disadvantages of low accuracy and slow speed in the active vision security inspection method,it is not suitable for real-time security inspection.Aiming at this problem,we propose an x-ray inspection image classification algorithm combining octave convolution(OctConv)with attention-based bidirectional Gate Recurrent Unit(GRU).Firstly,OctConv is introduced to replace the traditional convolution operation to divide the input feature vector into high and low frequency,and reduce the resolution of low frequency features,effectively extracting the features of security image and reducing the spatial redundancy.Then,the feature weight can be adjusted by dynamic learning through attention-based bidirectional GRU to improve the classification accuracy of threat objects.Finally,a lot of experimental results on SIXRay dataset show that the classification accuracy,AUC value and PRE of 8000 test samples are 98.73%,91.39%and 85.44%,respectively,with a time of 36.80 seconds.Compared with the current mainstream model,the proposed algorithm can improve the performance and speed of threat objects recognition in X-ray security images.
作者
吴海滨
魏喜盈
王爱丽
岩堀祐之
WU Hai-bin;WEI Xi-ying;WANG Ai-li;YUJI Iwahori(College of Measurement–Control Technology and Communication Engineering,Harbin University of Science and Technology,Harbin 150080,China;Computer Science,Chubu University,Aichi 487-8501,Japan)
出处
《中国光学》
EI
CAS
CSCD
北大核心
2020年第5期1138-1146,共9页
Chinese Optics
基金
国家自然科学基金(No.61671190)。
