Abstract: Aiming to make more comprehensive use of pedestrian features, this study proposes a person re-identification model combining Attention and Batch Feature Erasure Network(ABFE-Net) to solve the problem of reduced recognition accuracy caused by the partial occlusion of images in person re-identification.First, the lightweight attention module is embedded into ResNet-50 to autonomously learn the weight of each channel, enhance the learning ability of network features by strengthening useful features, and suppress irrelevant features, thereby extracting more discriminative global features of pedestrians.Second, the batch feature erasure method is used for deep features, randomly erasure the same region of feature maps in a batch to focus on the remaining local fine-grained features.Finally, the two features are merged to obtain a more comprehensive representation of pedestrian features, and similarity measures are performed on them and sorted to obtain the result of person re-identification.On the Market1501 dataset, Rank-1 and mAP reach 94.4% and 85.9%, and on the DukeMTMC-reID dataset, Rank-1 and mAP reach 88.3% and 75.1%, respectively.Experiments results reveal that compared with HA-CNN, PCB, and other methods, the ABFE-Net model can reinforce the discrimination of pedestrian characteristics and improve the performance of person re-identification.

Key words: person re-identification, batch feature erasure, attention mechanism, residual network, metric learning

摘要： 在行人重识别过程中，图像局部遮挡会造成识别准确率下降。提出一种结合注意力和批特征擦除的网络（ABFE-Net）模型，旨在学习具有辨别力的全局特征和局部细粒度特征，提高图像局部遮挡条件下行人特征的表达能力。将轻量级注意力模块嵌入到ResNet-50中自主学习每个通道的权重，通过强化有用特征和抑制无关特征增强网络特征的学习能力，提取行人更具辨别力的全局特征。对于深层特征使用批特征擦除方法，随机擦除同一批次特征图的相同区域，使得网络关注剩余的局部细粒度特征。将两种特征融合得到更加全面的行人特征表示，对其进行相似性度量并排序，得到行人重识别的结果。实验结果表明，与HA-CNN、PCB等方法相比，ABFE-Net模型在Market1501和DukeMTMC-reID数据集上的Rank-1和mAP分别达到94.4%、85.9%和88.3%、75.1%，能够明显增强行人特征的辨别性，提高行人重识别效果。

关键词: 行人重识别, 批特征擦除, 注意力机制, 残差网络, 度量学习

CLC Number: 

• TP391.41