Finger-vein recognition using supervised deep learning methods is highly effective for biometric recognition and identity verification. However, developing a generalized recognition model for finger-vein patterns requires a substantial amount of labelled data, which is costly and time-consuming. Contrastive learning presents a promising alternative by learning representations from raw sensor data without human labels. In this paper, we propose CL4F-FV, a Constrastive Learning-based Finger-Vein recognition method using Frequency-mixup augmentation and time-Frequency Feature Fusion. Specifically, CL4F-FV utilizes frequency-mixup augmentation to generate synthetic samples by mixing features from different finger-vein images in the frequency domain and employs time-frequency feature fusion to combine time-domain and frequency-domain representations for extracting comprehensive features. Both techniques enhance data diversity and the model’s generalization ability. To assess the performance of CL4F-FV, we conduct extensive experiments on three finger-vein datasets in terms of representations, fine-tuning, ablation study, and complexity analysis. The experimental results demonstrate that CL4F-FV has the potential to advance finger-vein recognition, offering improved accuracy, robustness, and practicality for real-world applications.