Automated Truffle Crack Detection Using Deep Learning and Machine Learning

Automated quality control in truffle production requires accurate detection of truffle-specific cracks to ensure product integrity. This study aimed to develop a machine learning framework to classify truffle cracks versus other cracks, addressing the need for reliable, automated inspection in industrial settings. A dataset of 300 images (150 truffle cracks, 150 other cracks) was used, with 5-fold cross-validation (192 training, 48 validation images per fold) and a test set of 60 images (30 per class). Three models were evaluated: VGG16 via transfer learning, Support Vector Machine with MobileNetV2 features, and EfficientNetV2B0 with Test-Time Augmentation. Cross-validation results showed EfficientNetV2B0 achieved the highest mean accuracy (0.929 ± 0.039) and Area Under the Curve (0.975 ± 0.026), followed by Support Vector Machine (accuracy 0.887 ± 0.031, Area Under the Curve 0.952 ± 0.025) and VGG16 (accuracy 0.867 ± 0.055, Area Under the Curve 0.942 ± 0.031). On the test set, EfficientNetV2B0 outperformed others with an accuracy of 0.933, precision of 0.933, recall of 0.933, and Area Under the Curve of 0.990. Support Vector Machine achieved a test accuracy of 0.917, with recall improving to 0.967 at an adjusted threshold of -0.3, while VGG16 recorded a test accuracy of 0.833, with recall of 0.867 at a threshold of 0.3. Confusion matrix analysis confirmed EfficientNetV2B0’s balanced performance, correctly classifying 28 of 30 truffle cracks and 28 of 30 other cracks. These models, trained on accessible computational resources, demonstrated robust performance, with EfficientNetV2B0 offering the most effective solution for automated truffle crack detection, enabling high-precision quality assurance in truffle production. Limitations include a modest dataset size, suggesting areas for future improvement.