用于车辆重识别的部件耦合Transformer网络

Part Coupled Transformer Network for Vehicle Re-Identification

基于卷积神经网络的车辆重识别模型在执行卷积和池化操作时,不可避免地会出现全局感受野狭小和局部信息丢失的情况,当光照、视角和分辨率等发生剧烈变化时,导致车辆重识别的鲁棒性和精确性急剧下降.为此,提出了部件耦合Transformer的车辆重识别网络,通过堆叠部件耦合Transformer块来搭建重识别模型,每一个部件耦合Transformer块利用部件自适应嵌入模块提取区分性的局部特征和Transformer层提取鲁棒性的全局特征.首先,部件自适应嵌入模块按照位置和伸缩量动态划分和调整特征图,增强模型对局部部件信息的感知能力;其次,Transformer层中利用自注意力机制增强网络模型对全局特征的表示能力;最后,部件自适应嵌入模块和Transformer层之间的耦合关系促进全局和局部特征协同合作.在VeRi-776和VehicleID数据集上的实验结果表明,CMC@1/CMC@5分别达到0.970/0.988和0.865/0.985,优于对比模型.

When vehicle re-identification models based on convolutional neural network perform convolutional and pooling operation,it is inevitable that the global sensitivity field will be narrow and local information will be lost.When illumination,perspective and resolution change dramatically,the robustness and accuracy of vehicle re-identification will decline sharply.Therefore.the part coupled Transformer(PCT)network for vehicle re-identification is proposed,which stacks multiple PCT blocks to build a re-identification model.Each PCT block uses a part adaptation embedding(PAE)module to extract discriminative local features and employs a Transformer layer to extract robust global features.Firstly,the PAE that splits and adjusts dynamically feature map according to positions and scales,enhances the ability to capture local parts.Secondly,the self-attention mechanism of Transformer layers improves the representation of global features.Finally,the coupling between PAE module and Transformer layer results in an effective cooperation between global and local features.Experimental results on VeRi-776 datasets and VehicleID datasets show that CMC@1/CMC@5 of PCT reaches 0.970/0.988 and 0.865/0.985 respectively,which is superior to the comparison models.