Weight prediction of pigs from Kinect image data

Abstract

In the Netherlands, pig farming contributes to an €8 billion production value to the national GDP. This includes €5 billion in exports, as the industry’s 250% self sufficiency (as reported in 2015), allows for a large number of exports. With descrepancies between the Netherlands and larger growers in Europe such as Germany and producer giants in Asia, it is crucial for each producer to provide the highest quality product in the market, particularly when exports are involved. The amount of meat a pig is capable of providing can be calculated from the carcass weight, obtained traditionally by either using weighing scales or hand measurements. This system is taxing to a company with regard to manual labor and human resources. This is even more so the case for operations involving hundreds of thousands to millions of pigs. A large dataset of frame-by-frame images of pigs has been provided by Topigs Norsvin, a Netherlands-based company renowned for its solutions to perfecting pig genetics through smart breeding. An automatic subsampling algorithm is used to retain the best quality images for training and testing. The dataset is then further downsized to equate the weight classes, eliminating bias. This project considers both a classification and a regression approach for weight prediction, producing output as discrete weight group labels and as continuous values.

A combination of image properties using the Matlab regionprops function and dense SIFT image features were used to create the weight prediction system. A transfer learning experiment was also performed using AlexNet on binary masked images and tested alongside the other predictors. The classifier’s optimal combination of predictors for estimating pig weight was found to be the SIFTs and CNN layer output, which rendered the aforementioned experiment useful to this study. Results obtained from the system report a classification accuracy of ∼63.5%. The training set used for classification was also run through a k-fold validation with 10 folds, which resulted in a 63.1% accuracy. Studying the classifier revealed that a significant amount of the classification error resided among the weight groups in the middle of the distribution. The regression performed substantially better, reporting an average error of 3.34kg and RMSE of 3.66kg. The spread of error was also considerably more impartial across weight groups compared to classification.