Abstract: Abstract: Training a fine-grained image classification （FGIC） model with limited data is a great challenge， where subtle differences between categories may not be easily discernible. A common strategy is to utilize pre-trained network models to generate effective feature representations. However， when fine-tuning the pre-trained model using limited fine-grained data， the model often tends to extract less relevant features， which triggers the overfitting problem. To address the above issues， this paper designs an new FGIC method named SDA-Net under low-data conditions， which optimizes the feature learning process by fusing the spatial self-attention mechanism and the self-distillation technique， which can effectively mitigate the overfitting problem caused by data scarcity and improve the performance of deep neural networks in low-data environments. Specifically， SDA-Net improves the intra-class representation by introducing spatial self-attention to encode contextual information into local features. Meanwhile， a distillation branch is introduced and the distillation loss is used in the augmented input samples， which realizes the deep enhancement and transfer of knowledge within the network. A comprehensive evaluation on three fine-grained benchmark data shows that SDA-Net exhibits significant performance gains compared to both traditional fine-tuning methods and the current SOTA low-data FGIC strategy. In 3 scenarios with 10% low-data volume， relative accuracies are improved by 30%， 47%， and 29%， respectively， compared to standard ResNet-50， and by 15%， 28%， and 17%， respectively， compared to SOTA.

Key words: Key words: deep learning, fine-grained image classification, low-data learning, self-distillation, self-attention, data augmentation