Detection of sugar beet seed coating defects via deep learning

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• The global seed coating market is rapidly expanding, driven by innovations aimed at improving crop performance and agricultural sustainability.
• Coating defects significantly impact seed quality and germination, necessitating precise identification and classification methods.

Purpose of the Study:

• To categorize coated sugar beet seeds based on coating defects using the YOLO (You Only Look Once) algorithm.
• To develop an efficient and effective method for detecting and classifying seed coating imperfections.

Main Methods:

• Utilized a dataset of 2000 high-resolution RGB images of coated sugar beet seeds under controlled lighting conditions.
• Employed YOLOv10-N, YOLOv10-L, and YOLOv10-X models for classification, with an 80% training and 20% validation split.
• Classified seeds into categories: normal, broken, star-shaped, and adherent coatings.

Main Results:

• YOLOv10-X achieved the highest test accuracies: 93% for normal, 94% for broken, 94% for star-shaped, and 95% for adherent coatings.
• YOLOv10-N demonstrated the fastest inference times, ranging from 11.4 ms to 11.9 ms across defect types.
• The study confirms the efficacy of image processing for controlling operational condition effects on seed quality.

Conclusions:

• The YOLO algorithm provides a robust solution for automated classification of sugar beet seed coating defects.
• Accurate defect detection using image processing enhances seed quality control and boosts agricultural production efficiency.
• This technology offers a scalable approach to maintaining seed quality amidst varying operational conditions.