SP-YOLO: A Robust Sweet Potato Surface Defect Detector for Complex Postharvest Environments
Results in Engineering · 2026-11 · 114017
DOI: 10.1016/j.rineng.2026.114017 ↗Kurz gefasst
Fachartikel in der Zeitschrift Results in Engineering, 2026-11. Verfasst von Minqiu Kuang und 11 weiteren Autor:innen. Die Arbeit nennt 57 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Hunan Provincial Innovation Foundation for Postgraduate.
Abstract (Originalsprache)
Surface defects such as mechanical damage, breakage, and insect holes directly affect the commercial grading, storage quality, and subsequent processing of sweet potatoes after harvest. To address the challenges of substantial defect-scale variation, loss of fine-grained features, and inaccurate localization of small and irregular defects in complex postharvest environments, this study proposes SP-YOLO, an improved sweet potato surface defect detection model based on YOLOv11n. At the feature-extraction stage, MobileNetV4 is introduced to reconstruct the backbone network, providing an alternative multi-scale feature representation. At the feature-enhancement stage, the original C2PSA module is replaced by C2BRA, which incorporates Bi-Level Routing Attention to strengthen feature responses to defect-relevant regions, weak-texture patterns, and boundary details while suppressing interference from complex backgrounds. At the detection stage, a Localization Quality Estimation (LQE) mechanism is incorporated into the detection head to learn a score correction from bounding-box regression-distribution statistics and adjust candidate ranking. A series of ablation and comparative experiments was conducted on a sweet potato surface defect dataset containing three representative defect categories: damage, breakage, and insect holes. The results show that SP-YOLO achieves a Precision of 90.0%, a Recall of 63.6%, an mAP 50 of 75.1%, and an mAP 50–95 of 57.5%. Among all compared detection models, SP-YOLO achieves the highest Precision and mAP 50–95 . Compared with the baseline YOLOv11n, these two metrics increase by 9.1 and 6.0 percentage points, respectively. However, Recall and mAP 50 decrease by 17.8 and 7.6 percentage points, respectively, indicating reduced detection coverage despite the gains in Precision and mAP 50–95 . The proposed model contains 5.45 M parameters and requires 21.0 GFLOPs, while maintaining an inference speed of 200.8 FPS. The ablation experiments further show that different module combinations have distinct effects on detection performance, with the complete model achieving the highest Precision and mAP 50–95 among the evaluated ablation configurations. This study provides a visual detection method for sweet potato surface defects in complex postharvest environments and a perception basis for subsequent quality assessment.
Förderung
- Hunan Provincial Innovation Foundation for Postgraduate
Zitieren
APA
Kuang, M., Zhou, J., Wu, T., Li, X., Huang, Z., Liu, D., Xiang, Y., Wu, B., Lu, Y., Xie, F., Zhang, Y., & Li, X. (2026). SP-YOLO: A Robust Sweet Potato Surface Defect Detector for Complex Postharvest Environments. Results in Engineering, 114017. https://doi.org/10.1016/j.rineng.2026.114017
Vancouver
Kuang M, Zhou J, Wu T, Li X, Huang Z, Liu D, et al. SP-YOLO: A Robust Sweet Potato Surface Defect Detector for Complex Postharvest Environments. Results in Engineering. 2026:114017. doi:10.1016/j.rineng.2026.114017
BibTeX
@article{Kuang2026,
author = {Kuang, Minqiu and Zhou, Jixin and Wu, Tianqi and Li, Xiaojian and Huang, Zhouhao and Liu, Dawei and Xiang, Yang and Wu, Bei and Lu, Yuchen and Xie, Fangping and Zhang, Yuxuan and Li, Xu},
title = {SP-YOLO: A Robust Sweet Potato Surface Defect Detector for Complex Postharvest Environments},
journal = {Results in Engineering},
year = {2026},
pages = {114017},
doi = {10.1016/j.rineng.2026.114017}
}
EUVIMED