The task of vehicle recognition aims to identify vehicles in the environment, while vehicle re-identification focuses on accurately matching the same vehicle across different scenes. Vehicle re-identification methods often suffer from features misalignment and interference, which cannot get an effective feature representation. To address this challenge, this paper proposes a vehicle re-identification network based on dual-branch feature fusion and spatio-temporal attention (DBF-NET), which integrates the low-level and high-level features through a multi-scale approach. The network includes a multi-scale feature module, an attention module and a feature fusion module. First, a multi-scale feature module captures diverse scale features of the targeted vehicle to enhance identification accuracy. Second, to reduce the impact of background noise on low-level features, an attention module is introduced into the low-level feature branch to filter out irrelevant background details. Third, a feature fusion module combines the low-level and high-level features to optimize the feature representation of vehicles. The proposed method fully integrates low-level features with rich texture information and high-level features with stronger semantic information, while also filtering out irrelevant background features. It enhances the capability of feature expression and resolves the problem of poor feature representation caused by feature misalignment and interference. Experimental results on a public vehicle re-identification dataset demonstrate that our method achieves superior mAP and rank-k performance compared to most existing algorithms.