Fig 1.
Image acquisition device with top and front views.
Table 1.
Detailed information of the CRP dataset.
Table 2.
Camera specifications of acquisition devices.
Fig 2.
Visual differences in CRP images captured by different devices.
Fig 3.
Overall architecture of the proposed method.
Fig 4.
Illustration of object detection and bounding box refinement.
Fig 5.
CRP images with extracted key regions, illustrating the exocarp patch and albedo patch used for vintage-related feature analysis.
Table 3.
Network structure.
Table 4.
Performance comparison of different methods on CRP classification.
Table 5.
Ablation study results comparing different model configurations.
Fig 6.
Comparison of classification Acc. across different feature interaction layers.
Fig 7.
Boxplots illustrating the stability of classification Acc. across multiple independent training runs for different models.
Fig 8.
Confusion matrix visualization.
Fig 9.
Comparative results of cross-domain classification Acc. between direct transfer and meta-learning adaptation on Xiaomi and Vivo target domains.
Fig 10.
Grad-CAM visualization.
Fig 11.
3D t-SNE visualization.