Multi-Modal vehicle re-identification (Re-ID) faces challenges due to severe modality gaps, varying image quality under diverse conditions, and background bias. In this paper, we propose FAMNet, a novel framework that effectively enhances foreground features, extracts modality-invariant representations, and dynamically reweights spectral inputs. Our approach consists of three modules: (1) a Foreground Enhancement Module guided by SAM-generated masks to suppress background interference; (2) a Common Feature Extraction Module trained with a view-aware contrastive loss to ensure cross-modal semantic consistency; and (3) a Modality Weight Prediction Module leveraging classification confidence to assign quality-aware modality weights. Extensive experiments on two public multi-modal vehicle Re-ID benchmarks, RGBNT100 and MSVR310, demonstrate that FAMNet achieves state-of-the-art performance and robust generalization across challenging illumination and environmental conditions. The code is available at https://github.com/shua1q1nanha1/FAMNet .