Thermal imaging-based characterization of fruit bruising using deep learning: a cross-dataset evaluation study

Abstract

Thermal imaging has emerged as an effective non-destructive technique for the measurement and characterization of internal quality defects in agricultural products. In this study, two thermal image datasets for bruise detection in apples and tomatoes are developed under controlled acquisition conditions, and their potential for thermal-based quality assessment is systematically investigated. The datasets are constructed using an active thermography approach combined with a multi-angle imaging protocol to capture temperature variations associated with tissue degradation. To evaluate the discriminative capability of the acquired thermal patterns, transfer learning-based deep learning models, including ResNet-18, MobileNetV2, and EfficientNet-B0, are employed as computational analysis tools. High classification performance is achieved under within-dataset evaluation, with accuracy exceeding 90%, confirming that thermal responses provide reliable indicators for bruise detection. To further assess the robustness of the measurement framework, cross-dataset experiments are conducted by applying models trained on one fruit type to another. The results reveal a substantial decrease in classification accuracy, ranging from 52.1% to 72.3% for Tomato→Apple and from 63.7% to 77.5% for Apple→Tomato, indicating strong dataset dependency and sensitivity to variations in fruit structure and thermal characteristics. Confusion matrix and ROC analyses demonstrate reduced discriminative capability under domain shift conditions and highlight systematic misclassification patterns. An asymmetric generalization behavior is also observed, suggesting that certain datasets exhibit more diverse and transferable thermal feature representations. These findings emphasize the limitations of current data-driven approaches in handling variability in thermal imaging-based characterization tasks. The proposed datasets and evaluation framework provide a reproducible benchmark for fruit bruise measurement and analysis, contributing to more robust non-destructive quality assessment.