by Busov Igor | student of NSU
Methods of analyzing quantitative characteristics of wheat shoots based on visual or tactile
expert assessment have some disadvantages: subjectivity, labor input. An alternative
method was implemented in the work, which is based on the analysis of images by computer
vision methods.
The following methods and algorithms were used in the work: a convolutional neural
network of the Unet architecture, a modified tgi index, with a further search for the
maximum of the objective function using a genetic algorithm, depth-first search (DFS),
Dijkstra’s algorithm, Student’s criterion, Bartlett’s criterion, K²-test D ‘Agostino, MannWhitney test, Kruskal-Wallis test, logistic regression, RandomForest machine learning
algorithm, ResNet architecture convolutional neural network with 18, 50 and 101 layers.
The system created in this work was tested on the extraction of two quantitative
characteristics: ploidy and 1102/L-25 genotype.
For the task of image segmentation provided for determining ploidy, the results of the best
models on test samples are 0.8463870901428718 (IoU), for determining the genotype —
0.8999361294443262 (IoU). Statistically significant differences in the distributions of
descriptors extracted from images of shoots of different ploidy were almost never found, but
statistically significant differences were found almost everywhere in the distributions of
descriptors of shoots images of different genotypes. The results of the best classification
models on test samples for solving the problem of determining ploidy were as follows:
without division into time points — 0.6538461538461539 (accuracy), the first time point —
0.6153846153846154 (accuracy), the second time point — 0.7692307692307693 (accuracy),
the difference — 0.46153846153846156 (accuracy). The best genotype classification model,
1102/L-25, predicted the genotype without error.