Banca de DEFESA: ADRIELLE CARNEIRO DE ARAÚJO SANTOS
Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.STUDENT : ADRIELLE CARNEIRO DE ARAÚJO SANTOS
DATE: 20/02/2025
TIME: 08:00
LOCAL: Sala de seminários da Pós-Graduação - Departamento de Zootecnia
TITLE:
PREDICTION OF BODY WEIGHT CATEGORY OF LAYING CHICKENS USING DEPTH SENSING AND MACHINE LEARNING
KEY WORDS:
artificial intelligence; classification; computer vision; convolutional neural networks; depth
images.
PAGES: 54
BIG AREA: Ciências Agrárias
AREA: Zootecnia
SUBÁREA: Produção Animal
SPECIALTY: Manejo de Animais
SUMMARY:
Monitoring the body weight of laying birds is essential to ensure egg production capacity. However,
current monitoring methods have proved insufficient to meet the needs of the poultry sector, and
alternatives are needed to automate and facilitate this process. The aim was therefore to develop and
find a deep learning model for classifying depth images, using methods based on computer vision, in
order to estimate the body weight category of laying pullets. Ninety birds of the Dekalb White strain
aged between 7 and 15 weeks were weighed and classified into three body weight categories: light,
medium and heavy. Eighteen pullets per category were selected for filming with a depth camera and the
captures were subjected to image pre-processing techniques. Four models were used for machine
learning classification: Convolutional Neural Network (CNN), Visual Geometry Group 16-layer
Network (VGG16), Residual Network with 50 layers (ResNet50) and Convolutional Neural Network
Next (ConvNeXt). The models were tested for: accuracy, precision, recall, F-1 score, ROC AUC,
specificity, Matthews correlation coefficient and Kappa score. CNNs performed better than the other
architectures used in terms of accuracy (95.4%), precision (95.5%), recall (95.4%), F1 score (0.954),
ROC AUC (0.96), specificity (96.4%), Matthews correlation coefficient (MCC) (0.93) and Kappa score
(0.93) at 15 weeks. At 7 weeks it was also superior in accuracy (80.9%), precision (83.7%), recall
(80.9%), F1 score (0.81), ROC AUC (0.86), specificity (73%), MCC (0.726) and Kappa score (0.715).
The VGG16, ConvNeXt and ResNet50 models had F1 scores of 0.636, 0.625 and 0.30 at 15 weeks, and
0.577, 0.545 and 0.263 at 7 weeks, respectively. Thus, the CNNs were able to successfully estimate the
body weight category of birds at 7 and 15 weeks of age using depth images, allowing for the extraction
of three-dimensional characteristics and demonstrating great potential for practical applications.
COMMITTEE MEMBERS:
Presidente - LILIAN FRANCISCO ARANTES DE SOUZA
Interno - CARLOS BOA VIAGEM RABELLO
Externo à Instituição - DENIS HENRIQUE PINHEIRO SALVADEO