深度学习在现勘图像分类中的应用  

Application of deep learning in image classification of crime scene investigation

作　　者：白小军[1,2] 刘颖 申宇飞 周冬 BAI Xiaojun;LIU Ying;SHEN Yufei;ZHOU Dong(School of Computer Science and Engineering,Xi'an Technological University,Xi'an 710016,China;Key Laboratory of Electronic Information Application Technology for Scene Investigation,Ministry of Public Security,Xi'an 710121,China)

机构地区：[1]西安工业大学计算机科学与工程学院,陕西西安710016 [2]电子信息现场勘验应用技术公安部重点实验室,陕西西安710121

出　　处：《西安邮电大学学报》2018年第5期43-47,共5页Journal of Xi’an University of Posts and Telecommunications

基　　金：电子信息现场勘验应用技术公安部重点实验室开放课题(EISI2016006)

摘　　要：针对刑侦现勘图像分类的需求,研究了典型的卷积神经网络VGG网络和Res网络及其训练优化方法。分别使用VGG网络和Res网络执行图像分类任务,给出了网络搭建与训练的方法,并在刑侦现勘图库上进行了测试,结果表明,采用VGG网络和Res网络执行现勘图像分类,效果优于传统的方法。最后,引入金字塔池化层以改善图像特征的质量,分别对VGG网络和Res网络结构进行优化,测试结果证实优化后的网络分类准确率进一步提升。According to the requirement of image classification of criminal scene investigation,typical neural networks of VGG and Res as well as their training and optimizing are investigated.VGG network and Res network are used to execute image classification task,the procedure of network construction and training is introduced,and then they are tested on the scene investigation image dataset.Test result show that the performance of scene investigation image classification by VGG and Res network is better than that by traditional methods.In the end,the pyramid pooling layer is introduced to improve the quality of image feature,and some optimization on the structure of VGG and Res network can thus be made.Further tests prove that further improvement on accuracy of image classification can be achieved by this optimization.

关 键 词：现勘图像 图像分类 神经网络 金字塔池化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]