DEVELOPMENT OF COMPUTER VISION ALGORITHMS FOR DIGITAL PHENOTYPING OF PEAS

This research aims to overcome key limitations of traditional pea breeding, namely the lengthy variety development cycle and the subjectivity of manual phenotyping, by developing automated image analysis methods. The study compares three computer vision methods applied to peas: YOLO-based detection, semantic segmentation for recognizing plant elements in dry and green samples (using a proprietary digital phenotyping setup), and an original algorithm for detecting stem nodes by analyzing stem width. The detection method demonstrated low accuracy for plant parts. Semantic segmentation achieved 65 % accuracy for dry and 76 % for green plants. The node detection algorithm demonstrated 100 % accuracy. The developed software package enables objective assessment of key pea phenotypic traits. Further development of the system is aimed at integration with neural networks for determining leaf surface area and the number of productive nodes, which creates the basis for accelerated pea breeding.