Ensemble Algorithm using Transfer Learning for Sheep Breed Classification
Divyansh Agrawal, Sachin Minocha, Suyel Namasudra, Sathish Kumar
Abstract
Sheep fostering is an increasing trend due to the huge demand for sheep wool, milk and mutton meat throughout the world. The export market value for sheep meat in Australia alone is approx. USD 5.23 billion. While the wool market worldwide was estimated USD 35 billion at the end of 2019 and it is predicted to reach USD 46.07 billion by 2025. Different sheep breeds have distinct characteristics like the wool of Merino sheep is costlier than most of the wool varieties available in the market. Therefore, it becomes important to identify the sheep breed to recognize the higher value characteristic of the corresponding sheep. This is indeed possible with human expertise, but this task is tedious and prone to human error. Thus, there is a need to identify sheep breeds with an accurate precision rate. This study aims to classify the sheep in a farm into four classes indigenous to Oceania. This paper proposes an ensemble model of the ResNet50 (Residual Network 50) and VGG16 (Visual Graphics Group 16) architectures that gives an improved sheep breed classification due to a boost in the learning. The ensemble model has been compared with five state-of-the-art transfer learning models, i.e. ResNet50, VGG16, VGG19, InceptionV3 (Inception Version 3) and Xception based on accuracy, log loss, recall score, F1 score and precision rate. The results show the efficiency of the proposed scheme.