A promising approach towards precise animal weight monitoring using convolutional neural networks

Abstract

Accurate monitoring of the live weight of pigs provides important information about the health state, the daily gain, and the time point for marketing. However, manual weight determination is time-consuming and stressful for both stockman and pig. In order to overcome these problems, non-invasive weighing mechanisms have to be established. In this study, we present an approach for live weight determination based on convolutional neuronal networks applied solely on the depth images of pigs, without further feature extraction. Our data basis consists of >400 pigs, recorded at four weighing time points, ending up with a weight range between 20 and 133 kg. Training and testing on this data, we achieved a coefficient of determination R2>0.97. Our results reveal that providing solely the images and the related weight to the ConvNets is sufficient to reach an accurate weight prediction. Therefore, our study can be viewed as a preliminary work that confirms the ability of using a ConvNets for accurate weight determination at different life stages. With the aim of using them under usual housing conditions for pigs, we increase animal welfare by precise animal monitoring in the sense of precision livestock farming.