In recent years, biometric recognition technologies have become widely used in various fields due to their ability to leverage unique physical and behavioral characteristics of individuals. Among biometric methods, palmprint recognition is preferred for its high stability, low invasiveness, and easy data collection compared to other approaches. In palmprint recognition, the Region of Interest (ROI) is the central area containing key biometric features and plays a vital role in identifying different individuals. However, current recognition methods still face several challenges. Traditional methods are easily affected by lighting or hand position, making their performance unstable. On the other hand, deep learning methods need a large amount of labeled data for training, which raises costs and brings up concerns about security and privacy. To solve these problems, we propose a new end-to-end Few-shot learning method called Multi-State Attention Fusion Network. This method uses a hierarchical structure to extract and combine multiscale pixel-level features, allowing it to gather useful information even from limited data. These features are then trained using Episodic Training combined with Metric-based learning, improving the model’s ability to learn and adapt effectively with just a small dataset. Experiments on the Tongji Dataset show that our method achieves impressive performance compared to other models. Our code is available at: https://github.com/simplesfish/Multi-State-Attention-Fusion-Network .