Research On Automatic Counting And Weight Determination Method Of Pre-weaning Piglets Based On Depth Image

Pig farming is an important part of husbandry in China,and pork accounts for 60% of total meat production.In large-scale farms,the weight of pigs is one of the important indicators of production.The work of the farrowing stable mainly includes piglet counting and weighing,which has been completed by manpower for a long time.With the continuous expansion of pig farms,the work of the farrowing stable has become complicated,which efficiency is relatively low,and it is difficult to correct human errors and ensure biological safety.In response to this problem,this article studies how to use computer vision to automate this work to achieve the purpose of obtaining piglet body weight without contact,accurately estimate the growth curve of a single animal,evaluate the variability within and between groups,and directly reflect the growth status of piglets and other information.In this study,a piglet individual segmentation model based on RGB images of piglets and a model for piglet weight prediction based on depth images were developed.The model can estimate the body weight of piglets throughout the growth cycle in a non-contact way,and draw the corresponding growth curve of each piglet,providing data support for the subsequent analysis of factors affecting the weaning weight of piglets and the number of weaned piglets.By comparing the weight prediction accuracy of various models,the finetuned Xception is finally selected as the optimal weight prediction model.The main results are as follows:(1)In this study,437 RGB pictures of piglets were collected and processed,and the label information files were obtained after labeling.The Mask R-CNN instance segmentation model was constructed by using the Tensor Flow framework.The model was not affected by the feeding environment and light.The final AP value is 0.86.(2)Using image enhancement methods to achieve data set expansion by image transposition,adding noise,the experiment found that image enhancement can significantly reduce the overfitting of the model.Without image augmentation,the mean absolute error on the training set was 0.11 kg with a coefficient of determination of 0.99,and the mean absolute error on the test set was 0.45 kg with a coefficient of determination of 0.78.The average absolute error obtained on the training set using image augmentation is 0.35 kg,the coefficient of determination is 0.88,and the average absolute error on the test set is 0.36 kg,the coefficient of determination is 0.86,the use of image augmentation improves the model generalization ability.(3)The prediction effect of the weight prediction model in different scenarios was compared and analyzed.The average absolute error obtained by using the images of piglets in the turnover box was 0.35 kg,and the coefficient of determination was 0.85.The mean absolute error obtained using images of piglets in the cage was 0.47 kg with a coefficient of determination of 0.71.The images taken in the turnover box build the model to predict better.(4)The transfer learning method was used to compare and analyze the performance of Inception V3,Res Net101 and Xception in predicting piglet weight,and Xception’s prediction effect was better than the other two models.In the test,the mean absolute error of the model is 0.36 kg,and the R2 is 0.86,which realizes the non-contact prediction of the individual body weight of a single piglet and provides convenience for pig production.In summary,this study aims to solve the problem of complicated piglet counting and weighing in the farrowing room,and there are errors in manual weighing.By using piglet RGB images and depth images,a set of analysis procedures for automatically obtaining piglet weight and a weight prediction model are constructed.The coefficient of determination is high,which provides help for actual production.