基于冗余特征正则化的车辆重识别算法

Vehicle Re-Identification Algorithm Based on Redundant Features Regularization

基于深度学习的车辆重识别算法使用空间关系不明确的滤波器提取特征,这些互不独立的滤波器会导致特征提取互相依赖且冗余,阻碍模型寻找数据的潜在规律.为此,提出一种显式的基于冗余特征正则化的车辆重识别算法Res-GC(ResNet grouping convolution).利用残差分组卷积网络阻止特征相互适应,以获取带有正则特性的车辆特征.引入BNNeck(batch normalization neck)方法解决交叉熵损失函数和三元组损失函数适用的车辆特征空间不一致的问题,提升正则特征使用效率.在车辆重识别的公共数据集VeRi-776和VehicleID上,Res-GC算法的实验结果均优于现有模型,验证了算法的有效性.

The vehicle re-identification algorithm based on deep learning uses filters with ambiguous spatial relationships to extract features.These non-independent filters will cause feature extraction to be interdependent and redundant,and hinder the model from finding potential patters in the data.To this end,an explicit vehicle re-identification algorithm ResNet grouping convolution(Res-GC)based on redundant feature regularization is proposed in this article.The residual grouped convolutional network is used to prevent the mutual adaptation of features to obtain vehicle features with regular characteristics.The batch normalization neck(BNNeck)method is introduced to solve the problem of the inconsistency of the vehicle feature space applicable to the cross-entropy loss and the triple loss function,thereby improving the efficiency of the use of regular features.On the public data sets VeRi-776 and VehicleID for vehicle re-identification,the experimental results of the Res-GC are ahead of the existing models,thus verifying the effectiveness of the proposed algorithm.