基于随机遮挡辅助深度表征学习的车辆再辨识
Random occlusion assisted deep representation learning for vehicle re-identification

在车辆再辨识中, 如何通过车辆外观学习到具有强区分度和鲁棒性的表示特征是至关重要的。为此, 本文提出一种基于随机遮挡辅助深度表征学习的车辆再辨识算法以提高车辆再辨识的准确率。首先, 本文所提算法通过引进随机遮挡对原始训练图片在局部区域进行随机遮挡, 一定程度上模拟了现实中的一些遮挡情况，不仅增加了训练样本的数量, 而且新增遮挡样本对于网络模型来说属于困难样本, 能够防止网络模型过拟合, 使得网络模型具有更强的鲁棒性；其次, 本文所提算法通过构建孪生网络对原始图片和随机遮挡图片进行分类和验证联合优化学习。实验结果表明,在VeRi和VehicleID这两个数据库上,所提算法优于现有多种车辆再辨识方法。
In vehicle re-identification, it is very important to learn more discriminative representations from the vehicle appearance. Therefore, a random occlusion assisted deep representation learning based vehicle re-identification algorithm is proposed to improve vehicle re-identification performance in this paper. First, the proposed algorithm employs the random occlusion (RO) method to randomly occlude the original training images, which simulates some occlusion situations in real world to a certain degree. One one hand, this increases the number of training samples. On the other hand, the new random occluded training samples are difficult for the learned model, which can prevent the model from over-fitting and achieve higher robustness. Then, the joint identification and verification learning optimization are performed on training the original images and occluded images through the developed simaese network. Extensive experiments conducted on the VeRi and VehicleID datasets have demonstrated that the proposed algorithm outperforms multiple state-of-the-art vehicle re-identification methods.