This paper proposes a self-reconstruction network for fine-grained few-shot classification. The method addresses the issues of overfitting and feature misalignment in traditional metric-based few-shot learning. The self-reconstruction network introduces a self-reconstruction metric module to diversify query features and a restrained cross-entropy loss to avoid over-confident predictions. By combining these components, the proposed method effectively alleviates overfitting. Extensive experiments on five benchmark fine-grained datasets demonstrate that the method achieves state-of-the-art performance on both 5-way 1-shot and 5-way 5-shot classification tasks. The self-reconstruction network reconstructs query features from support features and from the query features themselves, enhancing feature diversity and improving generalization. The method also incorporates a restrained cross-entropy loss to prevent over-confident predictions. The proposed network is trained using a joint loss function that combines cross-entropy and restrained cross-entropy losses. The experiments show that the method outperforms existing methods in terms of classification accuracy and stability. The results indicate that the self-reconstruction and restrained cross-entropy loss are effective in mitigating overfitting and improving the performance of few-shot classification. The method is implemented using ResNet-12 and ResNet-18 backbones and is evaluated on several benchmark datasets. The computational cost of the method is relatively low, making it suitable for practical applications. The paper concludes that the proposed self-reconstruction network is an effective solution for fine-grained few-shot classification.This paper proposes a self-reconstruction network for fine-grained few-shot classification. The method addresses the issues of overfitting and feature misalignment in traditional metric-based few-shot learning. The self-reconstruction network introduces a self-reconstruction metric module to diversify query features and a restrained cross-entropy loss to avoid over-confident predictions. By combining these components, the proposed method effectively alleviates overfitting. Extensive experiments on five benchmark fine-grained datasets demonstrate that the method achieves state-of-the-art performance on both 5-way 1-shot and 5-way 5-shot classification tasks. The self-reconstruction network reconstructs query features from support features and from the query features themselves, enhancing feature diversity and improving generalization. The method also incorporates a restrained cross-entropy loss to prevent over-confident predictions. The proposed network is trained using a joint loss function that combines cross-entropy and restrained cross-entropy losses. The experiments show that the method outperforms existing methods in terms of classification accuracy and stability. The results indicate that the self-reconstruction and restrained cross-entropy loss are effective in mitigating overfitting and improving the performance of few-shot classification. The method is implemented using ResNet-12 and ResNet-18 backbones and is evaluated on several benchmark datasets. The computational cost of the method is relatively low, making it suitable for practical applications. The paper concludes that the proposed self-reconstruction network is an effective solution for fine-grained few-shot classification.