基于注意力机制与多尺度融合学习的车辆重识别方法

Vehicle re-identification methods based on attention mechanism and multi-scale fusion learning

在车辆重识别任务中,通常会出现相机角度变化和场景变化等情况,导致重识别准确率降低,为此提出了一种基于注意力与多尺度融合学习的车辆重识别方法,在多尺度下提取并融合浅层细节信息和深层语义信息。首先,构造一种深度学习网络,通过注意力机制学习车辆图像的显著性特征;然后,在多个尺度下对描述车辆身份的信息进行提取,将浅层表达的细节信息和深层表达的语义信息相融合构造空间特征;其次,对空间特征进行分解与重组,得到具有空间鲁棒性的局部特征,并与全局特征融合,构造车辆身份重识别特征;最后,利用该特征计算不同车辆图像间相似度,判断是否具有相同的身份。实验结果表明:在VeRi-776数据集上测试得到的Rank-1指标达到了94.0%,mAP指标达到了72.2%,表明该方法在相机角度变化、场景变化等情况下可以有效提高车辆重识别的准确率。

In the task of vehicle re-identification,camera angle changes and scene changes often occur,which leads to the decrease of re-identification accuracy.Therefore,a vehicle re-identification method based on attention mechanism and multi-scale fusion learning was proposed in this paper.Shallow detail information and deep semantic information were extracted and integrated at multiple scales.Firstly,a deep learning network was constructed to learn the salient features of vehicle images by the attention mechanism.Then,the information describing the identity of vehicles was extracted at multi-scale level,and then the detailed information expressed at the shallow layers and the semantic information expressed in the deep layers were integrated to construct spatial features.Secondly,the spatial features were decomposed and reorganized to obtain local features of spatial robustness,which were merged with global features to construct vehicle identity re-identification features.Finally,these features were utilized to calculate the similarity between different vehicle images to determine whether they are of the same identity.The experimental results showed that the Rank-1 index tested on the VeRi-776 dataset reached 94.0%,and the mAP index reached 72.2%,revealing that our method is able to effectively improve the accuracy of vehicle re-identification in the case of camera angle changes and scene changes.