Individual sheep identification: image capture positioning analysis and classification models

• Grant number: 24/18074-5
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date until: December 01, 2024
• End date until: November 30, 2025
• Field of knowledge: Agronomical Sciences - Agricultural Engineering - Rural Buildings and Ambience
• Principal Investigator: Késia Oliveira da Silva Miranda
• Grantee: Nicole Morás Coutinho
• Host Institution: Escola Superior de Agricultura Luiz de Queiroz (ESALQ). Universidade de São Paulo (USP). Piracicaba , SP, Brazil

Abstract

Sheep farming, an ancient and widely practiced activity, faces significant challenges in Brazil, such as individual identification and tracking, which are still conducted using traditional methods like ear tags and markings, causing stress, along with a lack of data on animal health. This project addresses the need for individual identification of sheep, aiming at the well-being and health of the animals. Deep learning approaches will be utilized for sheep identification, based on visual analyses, starting with an initial assessment of camera positioning and a comparative evaluation of identification models using Inception V3 and SqueezeNet. Data collection will take place within the extension group "Intensive Production System for Sheep and Goats" (SIPOC) at ESALQ-USP, using 30 Dorper x Santa Inês sheep with an average body weight of 30 ± 5 kg. The data will be collected from five different positions (front, rear, top, and left and right sides), capturing 100 images from each position per animal, totaling 500 photos per sheep and 15,000 for the complete experiment, in order to determine the most effective model. Results will be analyzed using the Orange Data Mining software and two distinct feature extractors, Inception V3 and SqueezeNet, employing four classifiers: K-Nearest Neighbors (kNN), Random Forest, Support Vector Machine (SVM), and Neural Network. Evaluation metrics will include AUC (Area Under the ROC Curve), accuracy (CA), F1 Score, precision (Prec), sensitivity (Recall), Matthews Correlation Coefficient (MCC), and the confusion matrix. In this context, it is expected that individual identification of the animals will be possible, allowing advancements in integrated computer vision systems for monitoring aspects such as animal health and well-being.