基于金字塔卷积和随机擦除的行人重识别研究

Research on Pyramid Convolution and Random Erasing for Person Re-identification

为了解决目前行人重识别算法特征提取能力弱、泛化能力差的问题,提出了基于金字塔卷积和随机擦除的行人重识别模型。首先以ResNet50作为基准网络,使用带有不同尺寸卷积核的金字塔卷积模块来提取行人图像多尺度的特征,并添加局部特征随机擦除分支来增强模型泛化能力,降低过拟合的风险。最后使用难样本三元组损失函数和标签平滑损失函数,对模型进行联合训练,通过在Market1501和DukeMTMC-reID数据集上进行消融实验,验证了该模型改进的合理性。在和其他主流行人重识别算法的对比实验中,该方法也表现出了较好的识别性能。

To solve the problems of weak feature extraction ability and poor generalization ability of person re-identification algorithm,a person re-identification model based on pyramid convolution and random erasing is proposed in this paper.Firstly,ResNet50 is used as the backbone network.Pyramid convolution modules with different size convolution kernels are used to extract multi-scale features of person images.Random erasing branch is added to enhance the generalization ability of the model and reduce the risk of over fitting.Finally,the model is jointly trained with the hard triplet loss function and the label smoothing loss function.The ablation experiments on Market1501 and DukeMTMC-reID data sets verify the rationality of the proposed model.Compared with other advanced person re-identification algorithms,the method in this paper also shows better recognition performance.