A Fine-Grained Long-Tail Distribution Food Image Classification Model with Attention Mechanism

The fine-grained food image classification is a challenging task due to the presence of long-tail distributions in food categories, which results in imbalanced datasets. This imbalance will lead to biased models that underperform on rare classes, thereby affecting the model’s robustness and accuracy. This paper proposes FL-FoodNet, a classification model designed to enhance fine-grained recognition in food image classification. FL-FoodNet integrates both channel attention and spatial attention mechanisms, which help the model focus on intrinsic features and precise locations within food images. Copping with the challenges posed by the long-tail distribution of categories, Focal Loss is added, allowing the model to focus on underrepresented classes and improving generalization across diverse food types. Experimental results demonstrate that FL-FoodNet achieves superior performance on the Food-101 dataset and UEC-Food 256 dataset, with Top 1 accuracy of 90.75% and 85.51%, respectively, and Top 5 accuracy of 98.95% and 96.13%, outperforming existing fine-grained image classification models.