PSI-16 Estimation of pigs live body weight from digital images using reference objects

Abstract

Abstract The live body weight (LBW) is an important parameter providing guidance for estimation of growth and feed conversion efficiency, body condition, presence of disease, and management of housing, nutrition and animal health in different life stages of livestock. This research study explores the possibility of developing a semi-automatic analytic system that estimates the LBW of pigs by applying machine learning methods that use approximated biometric measurements extracted from digital images acquired with consumer-level cameras in the presence of a reference object. Images corresponding to 12 pigs were sampled on two different dates 1 month apart and acquired using a consumer-level Motorola X4 mobile phone. The best 3 images for each pig were selected for each time point. Six measurements were extracted from each image using ImageJ. A total of 72 data points were analyzed using RStudio, and the generated correlation plot confirmed the positive correlations between the 6 predictors and LBW. Five machine learning (ML) methods were used to model the dependency between the measured parameters and LBW using WEKA. The Random Forest model outperformed all the other models and predicted LBW with the highest prediction accuracy (97%) and the lowest prediction error (MAE = 6.54), which could represent a good candidate for further studies. In conclusion, the semi-automatic image-based system is a promising approach combining machine learning, digital image analysis and manual scale extraction with the aid of reference objects of known size for accurate pig LBW estimation. The next stage of this study aims to integrate automatic image acquisition and image processing solutions in the current approach. This novel system will be cost- and time-efficient, and it can contribute to the development of intelligent solutions for scientific research and enhancing animal productivity in commercial pig farms.