The MEAL300 dataset, comprising 300 food ingredient classes, poses a significant challenge for image classification models. In this study, we explore various Convolutional Neural Network (CNN) architectures and optimization techniques to improve classification accuracy on MEAL300. We evaluate baseline CNN models and apply transfer learning with fine-tuning strategies to adapt pre-trained models to the dataset. By optimizing fine-tuning methodologies and incorporating regularization techniques, we achieve significant performance improvements. Our proposed model, a fine-tuned ConvNeXtXLarge, attains a state-of-the-art accuracy of 93.79%, outperforming other architectures such as MobileNetV3Large (85.29%) and EfficientNetV2B0 (83.56%). These findings demonstrate the effectiveness of transfer learning and fine-tuning for large-scale food ingredient classification and contribute to advancing automated food recognition systems with applications in nutrition tracking, food safety, and smart kitchen technologies.