Cargando…

Development of an Early Embryo Detection Methodology for Quail Eggs Using a Thermal Micro Camera and the YOLO Deep Learning Algorithm

Poultry production utilizes many available technologies in terms of farm-industry automation and sanitary control. However, there is a lack of robust techniques and affordable equipment for avian embryo detection and sexual segregation at the early stages. In this work, we aimed to evaluate the pote...

Descripción completa

Detalles Bibliográficos
Autores principales: Nakaguchi, Victor Massaki, Ahamed, Tofael
Formato: Online Artículo Texto
Lenguaje:English
Publicado: MDPI 2022
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9371013/
https://www.ncbi.nlm.nih.gov/pubmed/35957378
http://dx.doi.org/10.3390/s22155820
Descripción
Sumario:Poultry production utilizes many available technologies in terms of farm-industry automation and sanitary control. However, there is a lack of robust techniques and affordable equipment for avian embryo detection and sexual segregation at the early stages. In this work, we aimed to evaluate the potential use of thermal micro cameras for detecting embryos in quail eggs via thermal images during the first 168 h (7 days) of incubation. We propose a methodology to collect data during the incubation period. Additionally, to support the visual analysis, YOLO deep learning object detection algorithms were applied to detect unfertilized eggs; the results showed its potential to distinguish fertilized eggs from unfertilized eggs during the incubation period, after filtering radiometric images. We compared YOLOv4, YOLOv5 and SSD-MobileNet V2 trained models. The mAP@0.50 of the YOLOv4, YOLOv5 and SSD-MobileNet V2 was 98.62%, 99.5% and 91.8%, respectively. We also compared three testing datasets for different intervals of rotation of eggs, as our hypothesis was that fewer turning periods could improve the visualization of fertilized egg features, and applied three treatments: 1.5 h, 6 h, and 12 h. The results showed that turning eggs in different periods did not exhibit a linear relation, as the F1 Score for YOLOv4 of detection for the 12 h period was 0.569, that for the 6 h period was 0.404 and that for the 1.5 h period was 0.384. YOLOv5 F1 Scores for 12 h, 6 h and 1.5 h were 1, 0.545 and 0.386, respectively. SSD-MobileNet V2 performed F1 scores of 0.60 for 12 h, 0.22 for 6 h and 0 for 1.5 h turning periods.