Real-world image datasets, especially in domains like food classification, often suffer from class imbalance and limited data diversity, which can significantly impact model performance. To address these challenges, we enhance the YOLOv8m-cls model by integrating advanced data augmentation techniques, specifically CutMix and MixUp, to improve robustness and generalization. Our experimental results demonstrate that the augmented dataset significantly enhances model performance. Compared to training on the original dataset, the model trained with augmentation achieves a lower validation loss (0.57687 vs. 0.93444), higher Top-1 accuracy (87.17% vs. 85.51%), and higher Top-5 accuracy (96.67% vs. 95.84%). Although the training time increases due to the larger effective dataset size, the improvements in generalization and accuracy justify the additional computational cost. This work underscores the importance of combining state-of-the-art models with tailored augmentation strategies to address the inherent challenges of real-world image classification tasks.