基于跨域双分支对抗网络车辆重识别策略

Vehicle Re-Identification Strategy Based on Cross Domain Dual Branch Adversarial Networks

为了减轻域偏差,提升算法的应用泛化能力,提出了一种基于跨域双分支对抗网络车辆重识别策略。首先充分挖掘源域的标记数据以适应目标域来缩小跨域偏差,并提出了一个名为双分支对抗网络的图像到图像的转换网络,从而有效保留来自源域的图像的属性。另外提出了一种结构注意力机制的特征学习模型,从而在抑制背景的同时提取显著特征。最后通过两个车辆重识别数据集试验结果证明提出的方法能够实现较高精度的车辆重识别效果,并且具有较好的泛化能力。

In order to reduce the domain bias and improve the application generalization ability of the algorithm,a vehicle re-identification strategy based on cross domain dual branch adversarial networks was proposed.Firstly,the marked data of the source domain were fully mined to adapt to the target domain to reduce the cross domain bias,and an image to image conversion network called dual branch confrontation network was proposed to effectively retain the attributes of images from the source do-main.In addition,a feature learning model based on structural attention mechanism was proposed to extract salient features while suppressing background.Finally,the experimental results of two vehicle re-identification data sets show that the proposed method can achieve high precision vehicle re-identification effect,and has good generalization ability.