An attempt to predict feed intake of dairy cows using RGB images

Monitoring of individual feed intake of cows is important for production efficiency. There are growing interests in using computer vision methods for measuring intake. As most developed models were 3D image-based, this study aims to predict feed intake using RGB images to capture the eating patterns of individual cows. Images were collected from multiple cows housed in a tie-stall barn. An RGB camera was positioned above the feeding plate, equipped with a weighing scale, that provides the ground truth of feed intake. Cows were fed twice daily, creating two feeding sessions per day. Data from 48 feeding sessions were collected, and images were selected at 10-minute intervals for each session to establish the time series dataset for training and evaluation of the model. While experimenting with multiple model architectures we achieved the best performance using transfer-learning on a pretrained EfficientNet model. We achieved an RMSE score of 3.58 ± 0.14 kg, MAE score of 2.97 ± 0.12 kg and an RMSPE of 22.7%. Although the outcomes are reasonable, important insights have been obtained concerning the limitations inherent in its present configuration. Our approach demonstrates significant potential for improved performance, which will be further investigated.