<![CDATA[Egg damage detection is important for maintaining food safety and operational efficiency in poultry production. Manual inspection is often inconsistent and difficult to scale, creating a need for accurate and computationally efficient automated methods. This study aims to develop and evaluate a transfer learning framework for automated egg damage classification using pretrained EfficientNet-B0. Three configurations were compared, namely Frozen Backbone Standard CE, Full Fine-Tuning Standard CE, and Full Fine-Tuning Weighted CE. The dataset was divided using a stratified 70:15:15 split, and the models were trained with AdamW, differential learning rates, early stopping, and image augmentation. Performance was assessed using accuracy, balanced accuracy, macro precision, macro recall, macro F1-score, weighted F1-score, confusion matrices, inference time, and Gradient-weighted Class Activation Mapping. The results show that Full Fine-Tuning Standard CE achieved the best overall performance, with an accuracy of 0.9917, balanced accuracy of 0.9947, macro F1-score of 0.9876, and weighted F1-score of 0.9917. The model correctly classified all non-damaged eggs and misclassified only one damaged egg, while maintaining an inference time of 29.53 ms per image. Weighted cross-entropy did not improve overall performance and increased false-negative predictions. Grad-CAM analysis showed that the model generally focused on relevant shell defects, although several errors were associated with background and contextual features. These findings demonstrate that full fine-tuning of pretrained EfficientNet-B0 provides an accurate, efficient, and interpretable solution for automated egg damage classification.]]>
Copyrights © 2026
