Title: Attention-based mechanism and feature fusion network for person re-identification

Authors: Mingshou An; Yunchuan He; Hye-Youn Lim; Dae-Seong Kang

Addresses: School of Computer Science and Engineering, Xi'an Technological University, Xi'an, China ' School of Computer Science and Engineering, Xi'an Technological University, Xi'an, China ' Department of Electronics Engineering, Dong-A University, Busan, South Korea ' Department of Electronics Engineering, Dong-A University, Busan, South Korea

Abstract: For the problem that person features cannot be sufficiently extracted in person re-identification, a person re-identification model based on attention mechanism is proposed. Firstly, person features are extracted using a hybrid network combining transformer's core multi-headed self-attentive module with the convolutional neural network ResNet50-IBN-a. Secondly, an efficient channel attention mechanism ECANet is embedded to make the model of this paper more focused on the key information in the person foreground. Finally, fusing the mid-level and high-level features in the model can avoid some discriminative features loss. The experimental results show that the provide model achieves 94.8% rank-1 and 84.5% mAP on the Market-1501 dataset; achieves 84.9% rank-1 and 65.9% mAP on the DukeMTMC-reID dataset; and achieves 40.3% rank-1 and 33.3% mAP on the Occluded-Duke MTMC dataset. Our proposed model compares well with some of the existing person re-identification models on these datasets mentioned above.

Keywords: attention mechanism; person re-identification; feature fusion; convolutional neural network; occluded person detection.

DOI: 10.1504/IJWGS.2024.137566

International Journal of Web and Grid Services, 2024 Vol.20 No.1, pp.74 - 92

Received: 04 Sep 2023
Accepted: 13 Dec 2023
Published online: 25 Mar 2024 *