Few-shot classification endeavors to recognize a query sample by leveraging a limited amount of support data, with prototype classifiers being frequently applied. While the prototype classifier is simple and non-parametric, it fails to fully leverage the prior information from the support samples, resulting in prototype bias. To address this, we introduce the Multi-stage Bayesian Prototype Refinement with Feature Weighting (MBPRFW). In our approach, we begin by implementing a feature weighting module to adjust the influence of each support sample. The adjusted features are then used as prior information to build the Bayesian prototype classifier, enabling the model to place greater emphasis on the most important aspects of the data. Ultimately, we incorporate a multi-stage inference strategy, in which the most distant support samples are filtered out at each stage. Through sample filtering, prototype representations and their associated classification scores undergo systematic recalibration. Therefore, we implement multi-stage Bayesian inference to effectively optimize conventional prototype classifiers. Comprehensive experiments corroborate the effectiveness of our strategy, demonstrating substantial improvement of the model’s discriminative capability in few-shot classification scenarios. The source code of this study is available at: https://github.com/CharlesXu2004/MBPRFW.