Vehicle re-identification involves utilizing robust vehicle query feature embeddings to obtain retrieval results in large-scale scenarios. Prior works have predominantly relied on CNN architectures, yet certain issues persist, such as loss of details and limited receptive fields due to pooling and convolution operations. To address these issues, this paper introduces a novel method that leverages multi-level feature semantic fusion and enhancement to overcome the limitations commonly associated with convolutional networks. We propose a multi-level hierarchical hybrid network architecture that integrates a Multi-Head Cross-Attention Feature Fusion module (CAF) and a Transformer-based Global Feature Enhancement module (GFE). Comparative analysis shows that our hybrid model surpasses traditional multi-branch CNN and Transformer-based architectures in terms of performance, establishing its efficacy in vehicle re-identification tasks.