Fine-Grained Image Classification Based on Self-Distillation and RFCAConv

To capture the details of fine-grained images accurately and address the challenges of significant intra-class variations and subtle inter-class similarities, a fine-grained image classification model based on self-knowledge distillation and receptive field coordinated attention convolution is proposed. This model aims to improve the classification accuracy of fine-grained images with subtle differences. The ResNet34 neural network is employed as the feature extraction backbone to extract features from fine-grained images. A self-distillation module is introduced to distill knowledge to soften hard targets, and the model is trained again with previously acquired knowledge to improve the sustainable utilization of meaningful knowledge. Besides, the receptive field coordinated attention convolutional module is utilized to replace the standard convolution operation to filters out redundant feature information and focuses on critical regions to distinguish fine-grained images, which further improve the performance of network. Furthermore, a dynamic temperature hyperparameter is indicated to achieve a better distillation performance. Finally, the effectiveness of the proposed model is validated on 3 benchmark datasets, FGVC-Aircraft, Stanford_Cars, and Stanford_Dogs, with classification accuracies of 96.29%, 98.81 %, and 99.71%, respectively. The expected calibration errors (ECE) can reach 1.25, 1.46, and 1.54, and the areas under the receiver operator characteristic (AURC) curves can reach 16.71, 17.07, and 17.65, respectively.