Detection of sugar beet seed coating defects via deep learning.

Abstract

The global seed coating market is expected to experience substantial growth, increasing from a 2023 valuation of USD 2.0 billion to an estimated value of USD 3.1 billion by 2028. This growth surge is primarily due to the consistent introduction of innovative seed coating technologies and formulations, which are designed to enhance seed quality, improve crop performance, and prioritize sustainability in agriculture. For this reason, the goal of this work is to categorize coated sugar beet seeds based on coating defects using the YOLO (You Only Look Once) algorithm. Coating defects can have a substantial impact on seed quality and germination rates; thus, seeds must be carefully identified and classified. Using the YOLO algorithm, it is possible to detect and categorize coating defects on sugar beet seeds, thereby enhancing seed quality and production swiftly, and effectively. To this end, totally high-resolution (3000 × 4000 pixel) RGB images of 2000 coated sugar beet seeds were used, which were obtained from a top-side open shooting box under constant 1150 lx daylight conditions to create an original database. The classification was performed on sugar beet seeds with normal, broken, star-shaped, and adherent coatings, based on 80% training and 20% validation rates with the YOLOv10-N, YOLOv10-L, and YOLOv10-X models. According to evaluations, the best test accuracies were obtained from YOLOv10X, 93% for normal coating, 94% for broken coating, 94% for star-shaped coating, and 95% for adherent coating. Additionally, the best inference times were obtained from YOLOv10N: 11.5 ms for normal coating, 11.7 ms for broken coating, 11.4 ms for star-shaped coating, and 11.9 ms for adherent coating. Therefore, it is possible that the negative effects of changing operating conditions can be brought into full control with image processing technologies.