Time-Sensitive Bruise Detection in Plums Using PlmNet with Transfer Learning
Yonis Gulzar, Zeynep Ünal
Abstract
This study introduces PlmNet, a novel CNN model specifically designed for time-sensitive bruise detection in plums using Near-Infrared NIR imaging. PlmNet integrates advanced architectural enhancements, including Global Average Pooling, dense layers, Batch Normalization, and Dropout layers, to optimize feature extraction and classification performance. The model was trained using the Adam optimizer with a learning rate of 0.0001 and a batch size of 8. A custom plum bruise detection dataset was developed, comprising three classes: Healthy plums, bruised plums captured 12 hours after bruising (12H_Bruised), and bruised plums captured 72 hours after bruising (72H_Bruised). PlmNet achieved a validation accuracy of 97.58%, a test accuracy of 97.17%, and consistently high precision (weighted average: 97.17%), recall (97.17%), and F1-score (97.17%) across all classes. For the challenging "12H_Bruised" class, the model achieved precision and recall exceeding 95%, significantly outperforming state-of-the-art model such as EfficientNetB3. These results demonstrate PlmNet’s robustness, efficiency, and potential for automating fruit quality assessment and bruise detection in agriculture, offering practical applications in supply chain management and quality monitoring.