基于深度学习的管制物品自动检测算法研究  

Automatic Detection Algorithm for Controlled Items Based on Deep Learning

作　　者：吉祥凌 吴军[1] 易见兵[1] 张晓光 Ji Xiangling;Wu Jun;Yi Jianbing;Zhang Xiaoguang(College of Information Engineering, Jiangxi University of Science and Technology, Ganzhou, Jiangxi 341000, China;Shayighai Eastimage Equipment Co. , Ltd. , Shanghai, 200120, China)

机构地区：[1]江西理工大学信息工程学院,江西赣州341000 [2]上海英迈吉东影图像设备有限公司,上海200120

出　　处：《激光与光电子学进展》2019年第18期68-78,共11页Laser & Optoelectronics Progress

基　　金：国家重点研发计划(2016YFC0800500);江西省自然科学基金(20181BAB202004);江西省教育厅科技项目(GJJ170515);江西理工大学科研基金(NSFJ2015-K14)

摘　　要：提出一种对图片分区域检测的特征融合目标检测算法。利用角度旋转方法对数据集进行扩增;在Single Shot MultiBoxDetector(SSD)算法的基础上,采用多尺度特征融合的方法在浅层特征图中融合更深层的特征,以扩大浅层特征图的感受野,提高小目标的检测精度;当输入图片较大时,如大于1024pixel×10^24pixel,对目标图像进行分区域检测。为了验证该算法的精度,选择VOC2007+2012通用数据集和SDCI2018管制物品数据集对所提算法的精度进行测试。结果表明:所提算法在VOC2007+2012通用数据集上的检测精度为80.3%,比SSD算法提高了1.4%;在SDCI2018管制物品数据集上的检测精度为97.9%,比SSD算法提高了2.2%。所提算法能够实时准确地检测出安检图片中的管制物品,特别是对于大图片中的小目标检测效果较好。This paper proposes a feature-fused object detection algorithm for image sub-region detection. First, a multi-angle rotation method is used to amplify the dataset. Then, based on the Single-Shot MultiBox Detector (SSD) algorithm, the deep features are fused to shallow features using the method of multi-scale feature fusion to enlarge the receptive field of the shallow feature map and improve the detection accuracy for small objects. If the input image is large, i. e., more than 1024 pixel ×1024 pixel, it is detected in separated regions. To verify the performance of the algorithm, the VOC2007 + 2012 and SDCI2018 datasets are selected to test the algorithm. The detection accuracy of the proposed algorithm on the VOC2007 + 2012 dataset is 80.3%, which is 1.4% higher than that of the SSD algorithm. The detection accuracy on SDCI2018 dataset is 97.9%, which is 2.2% higher than that of the SSD algorithm. The proposed algorithm can accurately detect the controlled items in security check images in real time, particularly for the small objects in large images.

关 键 词：探测器 管制物品 自动检测 SSD 特征融合 小目标 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]