Few-shot class incremental learning (FSCIL) tackles the challenge of incorporating new classes into a pretrained model using only a small number of training samples while preserving previously learned knowledge. This task becomes especially challenging with fine-grained datasets, where classes share subtle visual similarities, making it difficult to establish clear decision boundaries between old and new categories. Existing FSCIL methods often fail to extract feature representations unaffected by background interference and struggle to differentiate between highly similar classes, leading to the misclassification of new classes as old ones. To address these limitations, we create masked images as auxiliary data for three existing fine-grained datasets which are incorporated into the training process to mitigate the background noise and facilitate further research on mitigating background interference. In addition, a novel Purification Enhancement Network (PE-Net) framework is proposed which leverages an attention mechanism and a background erasing module to enhance the network’s focus on objects and improve feature representation by contrasting original data with masked data. Additionally, we develop a discriminative distance module to enhance the model’s discriminative detection capacity, establishing a robust foundation for future class incremental tasks. Extensive experiments on these fine-grained datasets demonstrate that our method consistently outperforms state-of-the-art approaches under FSCIL settings.