Accepted_test

Using YOLO neural network for segmentation of contiguous wheat grains

by Dmitry R. Avzalov | Evgeniy G. Komyshev | Dmitry A. Afonnikov | Novosibirsk State University, Novosibirsk, Russia | Institute of Cytology and Genetics, SB RAS, Novosibirsk, Russia | Institute of Cytology and Genetics, SB RAS, Novosibirsk, Russia; Kurchatov Genomic Center of the Institute of Cytology and Genetics, SB RAS, Novosibirsk, Russia

Abstract ID: 514

Event: BGRS-abstracts

Sections: [Sym 6] Section “Analysis of plant images for solving phenotyping problems”

Counting grains manually is a labor-intensive and time-consuming task. Development of approaches based on image analysis by computer vision methods is relevant at the moment because of low cost and ease of use. Also it is important to accurately define grain boundaries for measuring the characteristics of individual grains

The study involved the selection of five genotypes from the ITMI collection, with attention to cover genotypes having different grain colors. The neural network architecture from the Yolo family of the eighth version was chosen to solve the instance segmentation problem. The model was trained for detection and segmentation of individual grains. The training sample size was increased using the image augmentation method. The training and model running parameters such as number of epochs, batch size, optimizer? learning rate, IoU, confidence were selected.

The quality of the model performance on the test sample was evaluated using metrics such as CR (correct ratio) and mAP (mean average precision). CR was used to assess the quality of counting and was equal to 98.3%. The mAP metrics were computed to evaluate both detection and segmentation quality. The metrics for detection have higher values compared to the metrics for segmentation.