Optimizing Dual Training Approaches for Goat Face Recognition

Abstract

Livestock management faces a significant challenge in ensuring effective traceability and monitoring of food-producing animals. Advances in human biometric technologies have prompted the use of face recognition technology for goat identification and verification. This research project aims to enhance goat face recognition accuracy through the utilization of two versions of labelled images and video frame images. The primary challenge lies in determining the optimal type of training data to use. Regular validation on diverse datasets encompassing various goat face recognition scenarios is crucial to ensure the model's generalization capabilities. Furthermore, the dataset utilized reflects the complexities associated with livestock surveillance, including diverse settings and lighting conditions, posing significant challenges to accurate goat detection and recognition. The objectives of this study are to develop a robust system capable of effectively addressing these challenges and to strike a balance between training data inclusion and model generalization. The methodology employed involves leveraging Roboflow to extract frames from video data, label the images, preprocess them, and apply augmentation techniques to enhance dataset diversity. Frames from test videos, initially treated as "unseen" data, have been pivotal in improving the model's recognition capabilities by exposing it to realistic conditions. The project's methodology highlights the dynamic nature of model development and refinement in addressing real-world challenges in livestock management. Overall, the project aims to contribute to the advancement of goat detection and recognition systems, with promising results expected in improving livestock management practices. Ongoing experimentation and adaptation of techniques, such as adjustments to model architecture and hyperparameters, are conducted to achieve this.