Using convolutional neural networks for determining the body condition scores of dairy goats

Body Condition Score (BCS) as an indicator of herd productivity in animal husbandry is a scoring method used to measure the fitness level of animals. However, scoring animals is a time-consuming and costly process that requires good knowledge and experience, the machine learning can be applied to make scoring more objective in a short time. For this purpose, in this thesis, it is aimed to determine BCS with Convolutional Neural Networks (CNN) models that learn from goat images through deep learning. In the study, the classification performances of some recognized CNN models were compared using the images taken from front, back and pelvic region of 111 goats of Alpine and Saanen breeds. The pre-trained networks such as AlexNet, VGGNet, DenseNet, GoogleNet, Inception V3, MobileNet and ResNet and some versions of these, which are widely used in applications, were used as the CNN models in the study. These networks were trained and tested on five datasets from images taken from goats: original (DS1), normalized (DS2), augmented (DS3), edge-segmented (DS4) and SMOTE augmented (DS5). VGG-11 has the test performance 80.00%, 81.02%, 81.66%, 84.34% and 87.53% for accuracy on DS1, DS2, DS3, DS4 and DS5 datasets, respectively, and 81.93%, 81.98%, 82.80%, 84.16% and 86.78% for F1 metrics on the same datasets. The SMOTE to balance and equalize the number of images in the minority classes has led to an important increase in the performance of all models. This result shows that SMOTE is an augmentation method that should be applied because it improves the performance in CNN models. In the training of models, a grid-search for the hyperparameters such as learning rate, loss function, etc. is also important to find for an optimal combination of hyperparameters. In this study, with the grid search performed on the VS5 dataset, test performance levels of 88.81% accuracy and 87.90% F1 were achieved with 0.0001 learning rate and NLLLoss loss function using VGG-11. The findings obtained in the study showed that the CNN model with VGG-11 network using NLLLoss loss function and 0.0001 learning rate, can be used successfully on SMOTE balanced goat images datasets.