Comparison and optimization of pig mass estimation models based on machine vision

Li Zhuo,Mao Taotao,Liu Tonghai,Teng Guanghui

DOI: https://doi.org/10.3969/j.issn.1002-6819.2015.02.022

2015-01-01

Abstract:Pig’s weight is an important index for farmers to monitor pig’s growth performance and health. Traditional weighting brings lots of stress to animals and stockmen due to manual operation. Pig weighting based on machine vision is a non-intrusive, fast and precise approach, for it can free the farmer from heavy operational labor. The weighting system precision is assured by the estimation model. A lot of estimation models are addressed in pig weighting based on machine vision by researchers and engineers. Both independent variables and modeling approaches would influence the accuracy of estimated weight. In present work, comparison and optimization of the models were conducted, and the best model was validated in the real farm. In the first experiment, four growing pigs were raised from 30 to 124 kg. The feed was suppliedad libitum, and the lighting was in a 12/12 h light/dark cycle. A machine vision system was assembled and installed with two parallel cameras, an RFID (radio frequency identification devices) reader and a PC for capturing live images of pigs automatically. Using the assembled system, the pigs’ back areas were measured. The head and tail of pig in each picture was cut off for pig’s back area calculation. Five indexes of pig body (body length, width, height, hip width, and hip height) were measured manually every day. Linear regression, power regression, quadratic regression, principal component regression and RBF (radial basis function) artificial neural network were used to establish estimation models using the 79 sets of data. Those models were compared using the remaining 97 sets of data. The second experiment was carried out in the real farm to validate the favorable model. Five body indexes of 24 adult pigs were measured three times manually. The results of experiment station showed that all the reestablished models were suitable for pig weight estimation with varied accuracies. Linear regression model based on body sizes was the best one with a correlation coefficient (R2) of 0.996, while the linear regression model of hip height had the least correlation. Principal component analysis was used to solve problem of collinearity among body sizes. Nonlinear regression was used to fit pig mass increasing tendency. The power regression model of the principal component fitted the increase of pig weight best and had the highest correlation coefficient (R2) of 0.994. The average relative error of estimation weights was 2.02% compared with the 97 sets of experiment data. The correlation coefficient (R2) and relative estimation error of individual pig were 0.970 and 2.26% respectively, better than the model using back pixel area. Furthermore, the obtained average absolute error was 2.08 kg which was less than that of the model combining area and height, which was 4.6 kg. The established estimation model of pig weight using five body indexes contained more three-dimensional information of the pig body than the model using only area and the model combining area and height. Through the model comparison using the data of experimental station and the validation in the real farm, it is proved that the power regression model of principal component is the desired one for pig weight estimation using machine vision technology.