Deep Learning for Delicious Data: Improving Food Classification with Fine-Tuned DenseNet-201

Abstract

Accurate classification of food images constitutes a critical component of advanced dietary monitoring, nutritional analysis, and innovative food management systems. Nevertheless, food recognition remains a formidable challenge due to substantial intra-class variations, inter-class similarities, inconsistent backgrounds, and diverse presentation styles. In response to these challenges, this study introduces a transfer learning-based deep learning framework utilizing the DenseNet201 architecture for comprehensive food image classification. The proposed model is trained and evaluated on the publicly accessible Food-101 dataset, comprising 101,000 food images categorized into 101 distinct classes. In this research, DenseNet-201 is fine-tuned by substituting its original classifier layers with custom fully connected layers designed for multi-class classification. Data augmentation techniques—including random rotation, zoom, shear, and horizontal flip—are applied during preprocessing to enhance model generalization and mitigate overfitting. Additionally, early stopping and dropout regularization are employed to ensure stable convergence during training. The dense connectivity mechanism of DenseNet201 facilitates efficient feature reuse. It enhances gradient flow, thereby enabling the model to learn rich hierarchical visual representations essential for distinguishing between visually similar food categories. The proposed DenseNet201-based model is comprehensively evaluated using performance metrics, including accuracy, confusion matrix, and classification report, and its performance is compared with traditional CNN-based architectures. Experimental results demonstrate a significant improvement in classification efficacy, confirming the robustness and effectiveness of the fine-tuned DenseNet201 for real-world food recognition tasks. The findings underscore the potential of the proposed framework to serve as a scalable and accurate solution for computer vision applications in smart dietary assessment, restaurant automation, and mobile health monitoring systems.