Weighing live sheep using computer vision techniques and regression machine learning

This research arose from the need to aggregate computer vision technology and machine learning in sheep weight control and facilitate the weighing process of animals in farms. The experiment was conducted to collect the images of the animals and their weights, and later, the annotations of the images were made, generating a mask image dataset. We selected the attribute extraction algorithms that extracted shape, size, and angles with k-curvature. With these extracted data, we used the stratified five-fold cross-validation. Also, we used eight machine learning techniques aimed at regression, and the result obtained when compared to the metric Adjusted R2 was the technique called Random Forest Regressor to obtain Adjusted R2 0.687 (±0.09) and MAE of 3.099 (±1.52) kilograms. By performing the ANOVA test to check if it is statistically relevant using the Adjusted R2 measure, we got a p-value of 0.00000807 (8.07e−06). The contribution of the work is sheep weight prediction in a non-invasive way using images. Therefore, the results achieved make it possible to measure the animal’s weight with an MAE of 3.099 kg.