MP-CBA-YOLO: a morphology-preserving class-balanced object detection method for small-sample imbalanced ginseng grade detection

Ginseng grade determination depends on multidimensional morphological evidence, including rhizome-head segments, main-root morphology, fibrous-root distribution, epidermal luster, and local defects. For a long time, grades have mainly been assigned through comprehensive judgment by experienced experts according to industry standards. Because high-grade samples are limited in real-world collection and the visual boundaries between adjacent grades are continuous, deep-learning-based ginseng grade detection faces the combined challenges of small sample size, class imbalance, weak fine-grained differences, and interference from background variation. To address these issues, a standard YOLO-format ginseng grade object detection dataset was constructed. It contains 679 high-resolution images covering four grades: premium, first, second, and ordinary. All samples were acquired in a standardized photographic environment, and grade labels were confirmed by at least two experienced experts according to the T/thrs-Jilin Province Authentic Medicinal Material Ginseng Grading Standard. In addition, a Morphology-Preserving Class-Balanced Augmentation method, MP-CBA, was designed as a domain-prior-constrained data-level strategy for analyzing minority-class expansion in morphology-sensitive ginseng grading. The method performs conservative augmentation only for the minority premium and first classes during training. Small-angle rotation, slight scaling, low-amplitude translation, horizontal flipping, and brightness-contrast perturbation are used to expand sample coverage, while the validation and test sets remain completely unchanged. After validation-set-based model selection, YOLO11s with an input size of 640 was selected as the final configuration. In the held-out test-set evaluation, the selected YOLO11s-640 model achieved Precision, Recall, mAP50, and mAP50-95 values of 0.782, 0.913, 0.907, and 0.836, respectively. Validation-set comparisons further showed that excessive class-balanced augmentation and the P2 high-resolution detection head did not improve localization quality, indicating that model selection should be based on validation-set evidence rather than repeated test-set comparison.