基于CNN的机场安检危险品自动识别研究  

Research on Automatic Recognition of Dangerous Goods in Airport Security Inspection Based on CNN

作　　者：高强[1] 潘俊[1] 洪锐锋[1] GAO Qiang;PAN Jun;HONG Rui-feng(Department of Computer,Guangzhou Civil Aviation College,Guangzhou 510403,China)

机构地区：[1]广州民航职业技术学院计算机系

出　　处：《计算机技术与发展》2019年第10期95-99,共5页Computer Technology and Development

基　　金：广东省高等教育教学改革项目(201401017);广州民航职业技术学院科学技术项目(2017X0205,2018X0202)

摘　　要：机场安检是民航安全飞行的重要保障。针对机场安检危险品人工识别工作量大、效率低、易疲劳误判及危险品图像数据集不均衡导致识别准确率低的问题,提出一种基于GAN数据增强的卷积神经网络危险品自动识别模型。首先利用GAN使危险品图像数据集均衡化,然后将图像输入由4个卷积层、1个全连接层构成的卷积神经网络模型进行训练,训练时引入随机失活优化技术,使模型得到更好的识别效果。在2017公安一所危险品图像数据集上的实验结果表明,经过均衡化处理后,模型的识别准确率达到90.7%,较均衡化之前提高了33.4%。另外,经过对比实验,模型的识别准确率分别比GoogLeNet、AlexNet、ResNet高出5.8%、7.2%和5.4%。该模型具有较高的识别准确率及较好的实时性,对提升机场安检智能化水平具有积极意义。Airport security inspection is an important guarantee for civil aviation safety flight.According to the problems of heavy workload,low efficiency,easy fatigue misjudgment with artificial recognition and low recognition accuracy caused by imbalance of dangerous goods image dataset in airport security inspection,we propose a convolution neural network automatic recognition model for dangerous goods based on oversampling.Firstly,the GAN is used to equalize the dataset of dangerous goods image,and then the image is inputted into the convolution neural network model composed of four convolution layers and one full-connection layer for training.The stochastic deactivation optimization technique is introduced in the training to get better recognition effect.The experimental results on a dangerous goods image dataset of public security in 2017 show that the recognition accuracy of the model can reach 90.7%after equalization,which is 33.4%higher than that before equalization.In addition,the recognition accuracy of the model is 5.8%,7.2%and 5.4%higher than that of GoogLeNet,AlexNet and ResNet respectively.The model has high recognition accuracy and great real-time performance,which is of positive significance to improve the level of airport security intelligence.

关 键 词：危险品 CNN 自动识别 不均衡 GAN 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]