Deep Learning-Based Intelligent Sorting of Potato Tubers and Mineral Impurities: System Development and Experimental...

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Postharvest potato sorting faces challenges with mineral impurities like soil clods and stones, impacting efficiency and accuracy.
• Existing sorting methods may lack the precision and stability required for effective impurity removal.

Purpose of the Study:

• To design and develop an intelligent sorting system for potato tubers and mineral impurities.
• To enhance sorting efficiency, accuracy, and operational stability in postharvest conditions.

Main Methods:

• Developed a task-adapted detection model (YOLOv10n-PB) using YOLOv10n, a PSA module, and dynamic blur augmentation.
• Integrated a programmable logic controller and pneumatic actuators for online identification and removal.
• Conducted comparative experiments and L25(53) orthogonal tests to optimize parameters like conveyor belt speed, material spacing, and classification threshold.

Main Results:

• The YOLOv10n-PB model achieved a mean Average Precision (mAP@0.5) of 98.9% on the test set.
• Conveyor belt speed was the most significant factor influencing sorting accuracy, followed by material spacing and classification threshold.
• Optimal parameters (0.2 m/s belt speed, 9 cm spacing, 0.4 threshold) resulted in 98.3% overall sorting accuracy and 1.7% potato tuber false rejection rate.

Conclusions:

• The proposed intelligent sorting system demonstrates feasibility for accurate and stable automatic sorting of potato tubers and mineral impurities.
• The system effectively addresses challenges posed by soil clods and stones in postharvest sorting.
• Optimized parameters and the YOLOv10n-PB model contribute to high performance in real-world operating conditions.