摘要
针对车辆重识别中提取特征鲁棒性不高的问题,本文提出基于Vision Transformer的车辆重识别方法。首先,利用注意力机制提出目标导向映射模块,并结合辅助信息嵌入模块,抑制由不同视角、相机拍摄及无效背景引入的噪声。其次,以Vision Transformer远距离建模能力为基础提出通道感知模块,通过并行设计模型能够同时获取图像块之间和图像通道之间的特征,在关注图像块之间关联的基础上,进一步构建通道之间的关联。最后,利用卷积神经网络的局部归纳偏置,将全局特征向量输入到卷积注入模块中进行细化,并与全局特征联合优化,以构建鲁棒性的车辆特征。为了验证提出方法的有效性,在Ve⁃Ri776、VehicleID和VeRi-Wild数据集上分别进行了实验验证。实验结果证明,本文的方法取得了良好的效果。
Aiming at solving the problem of low robustness of feature extraction in vehicle re-identification,this paper proposes a vehicle re-identification method based on Vision Transformer.Combined with the auxiliary embedding mod⁃ule,the object-guiding projection module is proposed by using the attention mechanism to suppress the noise caused by different viewpoints,camera shots,and redundant backgrounds.Besides,based on Vision Transformer's long-distance modeling capabilities,a channel-aware module is proposed to build the association between patches and channels through parallel connections,which captures the discriminative features between patches and channels simultaneously.Finally,according to the local inductive bias of convolutional neural network(CNN),the global feature vector is input into the convolution injection module to extract local features,and jointly optimized with global features to construct ro⁃bust vehicle features.In order to verify the effectiveness of the proposed method,experiments were carried out on the Ve⁃Ri776,VehicleID and VeRi-Wild datasets.The experimental results prove that the method has achieved good results.
作者
于洋
马浩伟
岑世欣
李扬
张梦泉
YU Yang;MA Haowei;CEN Shixin;LI Yang;ZHANG Mengquan(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China;Institute of Information Science,Tianjin Academy of Agricultural Sciences,Tianjin 300192,China)
出处
《河北工业大学学报》
CAS
2024年第4期40-50,共11页
Journal of Hebei University of Technology
基金
国家自然科学基金资助项目(62276088,62102129)。
作者简介
第一作者:于洋(1981—),男,副教授。;通信作者:李扬(1982—),女,助理研究员,taas_liyang@126.com。
