Authors:
Mateus Silva
1
;
Breno Felisberto
2
;
Mateus Batista
3
;
Andrea Bianchi
1
;
Servio Ribeiro
3
and
Ricardo Oliveira
1
Affiliations:
1
Computing Department, Universidade Federal de Ouro Preto, Ouro Preto, Brazil
;
2
General Biology Department, Universidade Federal de Viçosa, Viçosa, Brazil
;
3
Biology Department, Universidade Federal de Ouro Preto, Ouro Preto, Brazil
Keyword(s):
Convolutional Neural Networks, Ant Ecology, Population Distribution.
Abstract:
A relevant challenge to be tackled in ecology is comprehending collective insect behaviors. This understanding significantly impacts the understanding of nature, as some of these flocks are the most extensive cooperative units in nature. A part of the difficulty in tackling this challenge comes from reliable data sampling. This work presents a novel method to understand the quantities and distribution of ants in colonies based on convolutional neural networks. As this tool is unique, we created an application to create the marked dataset, created the first version of the dataset, and tested the solution with different backbones. Our results suggest that the proposed approach is feasible to solve the proposed issue. The average coefficient of determination R 2 with the ground truth counting was 0.9783 using the MobileNet as the backbone and 0.9792 using the EfficientNet V2B0 as the backbone. The global average for the semi-quantitive classification of each image region was 86% for the
MobileNet and 88% for the EfficientNet V2-B0. There was no statistically significant difference between both cases’ average and median errors. The coefficient of determination was close to the statistical significance threshold (p = 0.065). The application using the MobileNet as its backbone performed the task faster than the version using the EfficientNet V2-B0, with statistical significance (p < 0.05).
(More)