Discovering the hidden personality of lambs: Harnessing the power of Deep Convolutional Neural Networks (DCNNs) to predict temperament from facial images

Abstract

The objective of this study was to define a more practical and reliable alternative to manual temperament classification methods that rely on the behavioral responses of animals individually subjected to various tests. Specifically, this study evaluated the correlation between facial image information and temperament based on deep convolutional neural networks (DCNNs) to predict the temperament of lambs based on their facial images. In the first phase, the lambs were categorized as to their temperament based on data acquired from a behavioral test to establish a ground truth for the temperament of the lambs. This enabled us to train (70%), validate (20%), and test (10%) deep-learning models in the second phase based on facial images and the corresponding temperament labels derived from the behavioral test. The performance of a custom deep convolutional neural network (C-DCNN) was compared to that of pre-trained VGG19 and Xception models for image classification. The Xception model achieved a training accuracy of 81%, which indicated that it learned well the underlying patterns in the data; however, lower validation (0.75) and test (0.58) accuracies indicate that it overfit the training data and did not generalize well to new samples. The VGG19 model, produced lower training (0.59), validation (0.46), and test (0.34) accuracies, which indicated that it did not learn the underlying patterns in the data as well as the Xception model. Furthermore, its precision (0.47), recall (0.42), and F1 score (0.41) indicated that the model performed poorly in identifying the classes correctly. The C-DCNN produced a moderate accuracy of 60%, which indicated that the model was able to predict the temperament traits of lambs with an accuracy of 60%, which was better than random guessing (33% accuracy), and demonstrated the potential of this approach in assessing temperament. The C-DCNN precision (0.69), recall (0.61) and F1 score (0.63) indicated that it had a moderate ability to correctly identify positive cases; however, the small size of the original dataset remains a limitation of the study because it might have caused the suboptimal performance of the models. To validate this approach, further research is needed based on a larger and more diverse dataset. We will continue to investigate the potential of deep learning and computer vision to predict animal personality traits from facial images based on large, diverse datasets, which might lead to more efficient and objective methods for assessing animal temperament and improving animal welfare.