Exploring Unseen Characteristics of Artificial Neural Networks for Improving Front Teat Placement Trait

We unfold the potential characteristics of artificial neural networks for cattle front teat placement trait, which refers to the location of a cow’s teats on its udder, specifically the placement of the front teats. For this purpose, we explore four deep learning models, namely the ResNet50 model (RNET50), the EfficientNet B1 model (ENET), a smaller customized convolutional model (SCNN), and a bigger customized convolutional model (BCNN). These models have variaitons in the feature extraction process. Our work contributes to the adoption of machine learning in cattle teat placement which is still in its early stage. We performed comprehensive experimental analysis on the front teat placement dataset we collected and prepared. We used the performance metrics: mean absolute error and average mean absolute error. We also exploit different augmentation techniques to encode variations in the collected data. These quantitative assessments provide good insights into how different architectures articulate around the front teat placement trait by capturing non-linear relationships conditioned on augmented and non-augmented data.