基于多视图架构深度神经网络的图像威胁识别  

Image Threat Recognition Based on Multiple View Architecture Deep Neural Network

作　　者：叶晴昊 涂岱键 毕奇 秦飞巍[1] 葛瑞泉 白静[3] YE Qinghao;TU Daijian;BI Qi;QIN Feiwei;GE Ruiquan;BAI Jing(School of Computer Science,Hangzhou Dianzi University,Hangzhou 310018,China;School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China;School of Computer Science and Engineering,North Minzu University,Yinchuan 750021,China)

机构地区：[1]杭州电子科技大学计算机学院,杭州310018 [2]武汉大学遥感信息工程学院,武汉430079 [3]北方民族大学计算机科学与工程学院,银川750021

出　　处：《计算机工程》2020年第11期261-266,共6页Computer Engineering

基　　金：国家自然科学基金(61702146,61762003,61972121);国家级大学生创新创业训练计划项目(201810336023);浙江省认知医疗工程技术研究中心开放课题(2018KFJJ05)。

摘　　要：由于藏匿物体的大小、形状和位置未知,且样本类别不均衡,常用的深度学习方法存在误报率较高的问题。为此,构建一种基于多视图架构的深度卷积神经网络模型。通过残差连接卷积神经网络对特征进行提取,使用基于稠密连接的长短期记忆注意力模型模拟人类多角度观察,以强化威胁信息表达,并基于焦点损失函数优化网络,从而构成端到端的架构。在HD-AIT毫米波人体威胁扫描数据集上的测试结果表明,相比其他基线模型,该模型的准确率和召回率分别可达到0.997、0.999。When applied to security checks,common deep learning methods have to address the high false alarm rate caused by the unknown size,shape and location of hidden objects as well as unbalanced sample categories.To deal with the problem,this paper proposes a deep convolutional neural network model based on multiple view architecture.The model uses convolutional neural networks with residual connections to extract features.Then a Long Short Term Memory(LSTM)attention model based on dense connections is used to simulate the process of human observations from multiple perspectives to enhance the expression of threat-related information.At the same time,the network is optimized based on the focus loss function to form an end-to-end framework.The experimental results on HD-AIT millimeter-wave scaned human body threating dataset show that the proposed model increases the accuracy to 0.997 and recall rate to 0.999 compared with other baseline models.

关 键 词：威胁识别 毫米波图像 深度学习 图像识别 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]