融合多尺度和频域特征的目标身份识别技术  

Target Identification Technology Integrating Multiscale and Frequency Domain Features

作　　者：徐勤功 郭杜杜[1] 赵亮 周飞[1] XU Qingong;GUO Dudu;ZHAO Liang;ZHOU Fei(College of Mechanical Engineering,Xinjiang University,Urumqi 830049,China)

机构地区：[1]新疆大学机械工程学院,乌鲁木齐830049

出　　处：《火力与指挥控制》2024年第1期175-181,189,共8页Fire Control & Command Control

基　　金：新疆维吾尔自治区自然科学基金(2019D01C043);浙江省智能交通工程技术研究中心开放课题(2021ERCITZJ-KF05)。

摘　　要：拍摄设备、角度和光线的差异以及相似目标的干扰,给跨设备目标身份识别任务带来了严峻挑战。针对识别过程存在的类内差异和类间相似问题,提出了一种融合多尺度和频域特征的识别模型,在主干网络中加入注意力机制提高模型对高辨识特征的关注度;在分支网络中,设计了一种基于注意力的多尺度扩张融合模块对不同深度特征进行多粒度采样融合,增强网络的空间映射能力;在后处理阶段构造自学习的频域卷积模块,实现多尺度和频域特征的融合,利用频域信息提高度量相似目标的准确率。经过实验,算法在Veri776和VehicleID数据集的平均均值精度(mAP)和首次命中精度(Rank-1)分别获得了81.60%、97.20%和90.50%、85.30%,结果优于近些年主流方法,能够满足跨设备的多目标身份识别要求。Differences in shooting devices,angles,and lighting,as well as the interference from the similar targets,severe challenges to the task of cross-device object identification are posed.Aiming at the problems of intra-class differences and inter-class similarities in the identification process,an identifica-tion model that combines multi-scale and frequency domain features is proposed.An attention mechanism to the backbone network is added to improve the attention of the model to high-discrimination features.In the branch network,an attention-based multi-scale expansion and fusion module is designed to perform multi-granularity sampling fusion on different depth features to enhance the spatial mapping ability of the network.A self-learning frequency domain convolution module is constructed to realize the fusion of multi-scale and frequency domain features during the post-processing phase,and the frequency domain information is used to improve the accuracy of measuring the similar targets.After experiments,the aver-age mean accuracy(mAP)and first hit precision(Rank-1)of the algorithm in the Veri776 and Vehicle ID datasets are 81.60%,97.20%,90.50%,and 85.30%,respectively,and the results are better than those of the mainstream methods in recent years.And it can meet the requirements of multi-target identification of cross-equipment.

关 键 词：目标身份识别 机器视觉 频域卷积 多尺度融合 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]