摘要
为了满足涉烟经济犯罪研判工作的检测精度,本文提出一种基于深度学习的涉烟犯罪智能检测系统。该系统使用YOLOX-Cr算法对涉烟图像进行目标检测。通过引入坐标注意力用于聚合涉烟图像中的关键性特征;使用Transformer计算全局特征信息的远程依赖关系,从而增强网络对前景和背景的辨别能力。在构建的首个涉烟数据集Cr-12K上测试了系统的工作性能,结果表明:本文算法对涉烟图像检测的均值平均精度(mAP)达到87.0%,相较YOLOX提升了1.4%,填补了涉烟经济犯罪领域智能化检测技术的空白。
In order to meet the requirements of detection accuracy in the judgment of tobacco-related eco⁃nomic crimes,this paper proposed a tobacco-related target detection algorithm based on coordinate atten⁃tion and Transformer.Based on the original YOLOX algorithm,this algorithm introduced coordinate atten⁃tion to aggregate key features in tobacco-related images.It used Transformer to compute long-range de⁃pendencies of global feature information to enhance the network capability in discriminating foreground and background.The performance of the system was tested on the first tobacco-related data set Cr-12K,and the results showed that the mean average precision(mAP)of the YOLOX-Cr algorithm reaches 87.0%.Compared with the YOLOX algorithm,the detection accuracy of the algorithm was increased by 1.4 per⁃cents.This algorithm filled the gap of intelligent detection technology in the field of tobacco-related eco⁃nomic crimes.
作者
黄紫婷
赵歆波
王洁钊
王瑞麟
HUANG Ziting;ZHAO Xinbo;WANG Jiezhao;WANG Ruilin(School of Northwestern Polytechnical University,Ningbo Research Institute,Ningbo 315103,China)
出处
《中国体视学与图像分析》
2023年第1期86-97,共12页
Chinese Journal of Stereology and Image Analysis
基金
宁波市自然科学基金(202003N4367)
国家自然科学基金面上项目(61871326)