CAJ | 학술논문

基于机器视觉的猪体质量估测模型较多，但模型缺乏在实用性、准确性的对比，最佳模型没有定论。该文总结了已有的估测算法，基于79组背部图像面积、实际面积、体长、体宽、体高、臀宽、臀高数据，使用线性回归、幂回归、二次回归、主成分线性回归、RBF（radial basis function，径向基函数）神经网络等方法，重建了13种体质量估测模型，并比较了13种模型的估测精度。结果表明，基于体长、体宽、体高、臀宽和臀高的线性回归模型具有较好的估测精度，估测值与真值的相关系数达到了0.996。利用主成分法去掉体尺的共线性，利用曲线回归解决残差不均匀问题，更加符合猪体质量增长趋势，结果表明基于主成分的幂回归模型具有较高的相关系数和较低的标准估计误差，对于97组数据的估测平均相对误差为2.02%。使用猪场实测24组数据验证模型，估测质量与测量值相关系数为0.97，估测平均相对误差为2.26%，标准差为1.78%，优于基于面积和面积体高结合的估测模型，平均绝对误差为2.08 kg，优于面积体高结合方法的平均绝对误差。试验证明使用多个体尺的主成分幂回归体质量估测模型较为精确，可用于机器视觉估测猪体质量的应用中。
기우궤기시각적저체질량고측모형교다，단모형결핍재실용성、준학성적대비，최가모형몰유정론。해문총결료이유적고측산법，기우79조배부도상면적、실제면적、체장、체관、체고、둔관、둔고수거，사용선성회귀、멱회귀、이차회귀、주성분선성회귀、RBF（radial basis function，경향기함수）신경망락등방법，중건료13충체질량고측모형，병비교료13충모형적고측정도。결과표명，기우체장、체관、체고、둔관화둔고적선성회귀모형구유교호적고측정도，고측치여진치적상관계수체도료0.996。이용주성분법거도체척적공선성，이용곡선회귀해결잔차불균균문제，경가부합저체질량증장추세，결과표명기우주성분적멱회귀모형구유교고적상관계수화교저적표준고계오차，대우97조수거적고측평균상대오차위2.02%。사용저장실측24조수거험증모형，고측질량여측량치상관계수위0.97，고측평균상대오차위2.26%，표준차위1.78%，우우기우면적화면적체고결합적고측모형，평균절대오차위2.08 kg，우우면적체고결합방법적평균절대오차。시험증명사용다개체척적주성분멱회귀체질량고측모형교위정학，가용우궤기시각고측저체질량적응용중。
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.