基于深度学习和属性特征的行人再识别算法

Person re-identification algorithm based on deep learning and attribute features

随着基于神经网络的算法在图像领域的不断发展,神经网络算法在行人再识别领域也逐渐成为主流的算法。目前,大多数神经网络算法常把网络的最后一层特征用于行人分类,而很少关注网络中间层输出的特征。另一方面,行人属性特征作为一个有效的局部特征,是神经网络提取特征的一个重要补充。基于Resnet50网络,结合网络中间层特征和行人属性特征,提出了一个新的行人再识别算法。在Market-1501和DukeMTMC-reID数据集上进行实验,实验结果表明,所提算法相较于目前主要算法,识别准确率有较大的提升。

With the continuous development of neural network-based algorithms in the image field,neural networkbased algorithms have gradually become mainstream algorithms in the field of person re-identification.At present,most neural network algorithms often use the last layer of network features for pedestrian classification,and rarely pay attention to the features of the network′s middle layer output.On the other hand,as an effective local feature,the person attribute feature is an important supplement to the feature extracted by the neural network.Based on the Resnet50network,this paper proposes a new person re-identification algorithm based on the feature of the middle layer of the network and person attribute features.Experiments are performed on the Market-1501and DukeMTMC-reID datasets.The experimental results show that compared with the current main algorithms,the proposed algorithm improves the recognition accuracy greatly.