Automatic Detection of Walnut Kernel and Shell Using Computer Vision: A Comparative Study

Lee J.M.E.; Fiorela S.Y.; Hernan I.E.C.; Francisco A.A.R. · 2026 The 9th International Conference on Image and Graphics Processing, ICIGP 2026 (2026) · DOI: 10.1109/icigp68997.2026.11619657

Resumen

The separation of walnut kernels and shells in walnut processing remains challenging due to their similar morphology and color. Computer vision and deep learning provide an alternative to traditional mechanical and pneumatic methods, which have limitations in reliability and generalization. This study compares four architectures, YOLOv5, YOLOv8, YOLOv11, and U-Net, offering a reference to evaluate their performance in kernel and shell detection and segmentation and to determine which architecture is most suitable for this task. A custom dataset of 400 RGB images was used, acquired under controlled lighting and a uniform background, and annotated with bounding boxes and binary masks. Results indicate that YOLOv11n achieved the highest recall (99.55%), F1-score (99.48%), and mean Average Precision in the range from 0.5 to 0.95 (82.47%), although it exhibited lower tolerance to noise. YOLOv5n achieved similar metrics, demonstrated greater robustness to visual perturbations (up to 25%), and had the highest processing speed (2.70 FPS), suggesting it is the most suitable option for industrial real-time applications under variable conditions. U-Net also demonstrated good noise resilience, although with lower inference speed. These results provide objective criteria for selecting and adapting architectures in industrial walnut processing and in future field applications.

Instituciones: Universidad Peruana de Ciencias Aplicadas