<![CDATA[Food recognition systems are commonly developed under the assumption that training labels are fully accurate. In real-world applications, however, food image datasets frequently contain noisy annotations caused by incorrect user inputs, weak labeling mechanisms, or automated data collection processes. This study investigates the robustness of supervised food recognition under synthetic label noise using the Food-101 dataset Food-101. The research employs EfficientNet-B0 as a computationally efficient backbone model and compares conventional cross-entropy learning with a robust training approach that integrates two mechanisms: (1) small-loss sample reweighting to reduce the influence of potentially corrupted samples, and (2) an early-learning stopping strategy based on the memorization gap between noisy training accuracy and clean validation accuracy. Symmetric label noise levels of 20% and 40% are introduced only into the training data, while validation and testing datasets remain unaffected. Experimental results on a 20-class subset demonstrate that the proposed approach substantially improves clean test accuracy from 0.6476 to 0.8176 under 20% noise and from 0.5636 to 0.6928 under 40% noise. In addition, probability calibration performance measured by Expected Calibration Error (ECE) is improved from 0.1474 to 0.0813 at the 40% noise setting. Additional experiments on the complete 101-class dataset also reveal consistent performance improvements despite shorter training durations. The findings suggest that combining loss-aware sample weighting with memorization-aware early stopping can provide an efficient and practical solution for building robust and reliable food recognition models in noisy-label environments.]]>
Copyrights © 2025
