Abstract:

Most of the person re-identification algorithms combined with deep learning are mainly based on single branch network structure, and most of them use the global feature information of the picture, so it is easy to miss the key pedestrian information, which leads to the decline of metric learning effect and algorithm accuracy. Therefore, in order to make the network obtain more key pedestrian information and reduce the loss of obtaining pedestrian and detail information, the network learning of pedestrian characteristics is strengthened. Based on the network structure of resnet-50, the multi-branches network structure design is adopted. Considering the training difficulty and calculated amount, this paper chooses the structure which merging three independent branches that are consisted of random erasuring branch, global learning branch and local learning branch. On this basis, the adjustment and optimization are carried out according to the experimental data. Finally, the weight of loss function is allocated by combining the least square method to make the model more accurate. The experimental results show that the three branches structure are complementary. Compared with the basic single branch and multi-branch network, the algorithm improves the person re-identification algorithms accuracy.

Key words: pedestrian re-identification, multi-branch network, random erasing, local feature, deep learning

摘要：

大部分结合深度学习的行人重识别算法主要以单分支的网络结构为主，且大多利用图片的全局特征信息，这样易错失关键行人信息，导致度量学习效果、算法精度下降。因此，为使网络获取到更多的关键行人信息，减少对行人局部、细节信息的错失，加强网络对行人特征的学习。基于ResNet-50的骨干网络，采取多分支的网络结构设计，综合考虑训练难易、运算量，选择融合三个独立分支的结构设计：随机擦除分支、全局学习分支、局部学习分支，并在此基础上根据实验数据进行调整优化，最后再结合最小二乘法分配损失函数权重使模型更具鲁棒性，实验结果表明，三个分支具有互补性，使用融合分支特征做算法测试时，相比基础的单分支、多分支网络，该算法使得行人重识别精度提升。

关键词: 行人重识别, 多分支网络, 随机擦除, 局部特征, 深度学习