基于联合条纹关系的车辆重识别

Re-identification of vehicles based on joint stripe relations

为了解决车辆重识别过程中因车辆特征图分块所导致的空间信息丢失问题,提出一种联合条纹特征之间关系的模块以弥补丢失的空间信息。首先,针对车辆特殊的物理结构,构建了一种双分支神经网络模型,对输出的特征图进行水平和垂直均等分割并在不同的神经网络分支上进行训练;然后,设计多激活值模块以减少噪声并丰富特征图信息;接着,使用三元组和交叉熵损失函数对不同的特征进行监督训练以约束类内距离并扩大类间距离;最后,设计批量归一化(BN)模块消除不同损失函数在优化方向上存在的差异,从而加速模型的收敛。使用所提方法在VeRi-776和VehicleID两个公共数据集上进行实验,结果表明该方法的Rank1值优于现有最好的方法VehicleNet,验证了其有效性。

In order to solve the problem of spatial information loss caused by the splitting of vehicle feature maps in the process of vehicle re-identification,a module combining the relationship between stripe features was proposed to compensate for the lost spatial information.First,a two-branch neural network model was constructed for the special physical structure of the vehicle,and the output feature maps were divided horizontally and vertically equally and trained on different branches of the neural network.Then,a multi-activation value module was designed to reduce noise and enrich the feature map information.After that,triplet and cross-entropy loss functions were used to supervise the training of different features to restrict the intra-class distance and enlarge the inter-class distance.Finally,the Batch Normalization(BN)module was designed to eliminate the differences of different loss functions in the optimization direction,thereby accelerating the convergence of the model.Experimental results on two public datasets VeRi-776 and VehicleID show that the Rank1 value of the proposed method is better than that of the existing best method VehicleNet,which verifies the effectiveness of the proposed method.