基于SSD的多特征刑侦场景识别  

Multi-feature detective scene recognition based on SSD

作　　者：姚红革[1] 白小军[1,2] 杨浩琪 YAO Hongge;BAI Xiaojun;YANG Haoqi(School of Computer Science and engineering,Xian University of Technology,Xi＇an 710021;Electronic information Scene Inquisition application technology Ministry of Public Security Key laboratory,Xi＇an 710121)

机构地区：[1]西安工业大学计算机科学与工程学院,西安710021 [2]电子信息现场勘验应用技术公安部重点实验室,西安710121

出　　处：《西安邮电大学学报》2018年第4期28-33,共6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：将基于深度学习的SSD(Single Shot MultiBox Detector)目标检测方法应用于刑侦图像目标检测中。通过对目标进行多尺度特征提取,将小目标与大目标采用不同级别的特征图方式进行融合识别。实验测试结果表明,SSD方法明显地提高了小目标在刑侦图像中的检测率,且与Faster R-CNN相比发现,在置信阈度为0.5时,SSD的检测精度接近Faster R-CNN,mAP(Mean Average Precision)达到94.8%,检测速度远超Faster R-CNN,帧频FPS达到58Hz。实验结果说明SSD方法在刑侦图像目标识别上具有特别优势。The SSD （Single Shot Multi-Box Detector） object detection method based on Deep I.earning is applied to object detection in criminal investigation images in this paper. Based on multi-scale feature extraction of objects, the small object and the large object are fused through the feature map of different levels, and thus the detection rate of the small object in the criminal detection image can be improved. Experimental results are compared with those by Faster R-CNN method. When the confidence threshold is 0.5, the detection precision of SSD is close to that of Faster R-CNN with its mAP is 94. 8~. While its detection speed is far beyond that Faster R-CNN and FPS can reach with its FPS is 58. Experiments show that the SSD method has a special advantage object recognition of criminal detection images.

关 键 词：SSD目标检测 刑侦图像 小目标 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]