Research On Non-contact Measurement Method Of Pregnant Sows Backfat Thickness Based On Computer Vision

The reproductive performance of sows is an important factor in determining the economic benefits of pig farms.Backfat thickness(BF)is an important physical trait of pregnant sows and is closely related to their reproductive performance and service life.Controlling the BF of pregnant sows within a reasonable range is crucial to ensure production and extend their service life.Moreover,different gestation stages have different requirements for the optimal BF range,and dynamic monitoring of pregnant sows BF is needed to ensure pig production capacity.However,the current method of measuring pregnant sows BF in pig farms mainly uses ultrasonic measurement,which requires the breeder to measure each pregnant sow with an ultrasound instrument,leading to high labor intensity and stress response in the sows.Therefore,the use of computer vision technology for non-contact measurement of pregnant sows BF is an urgent need for large-scale pig farms and has broad application prospects.Based on this,computer vision technology was used to develop a non-contact measurement method for measuring pregnant sows BF.Different gestation stages of sows were taken as the research object to make an image-based dataset for BF and a non-contact measurement model for BF was built.The main work and conclusions are as follows:(1)An image data collection system was built,the image data of pregnant sows and their BF and body size data were collected to create a BF image dataset and a body size dataset of pregnant sows.An image data acquisition platform was designed and an image data acquisition program was developed.Video data of the back of 106 pregnant sows at different gestation stages were collected from overhead view using a self-built image acquisition platform.At the same time,their corresponding BF data,body length,hip height,hip width,and other body dimension data were collected by hand.Finally,the video data and the manually collected data were pre-processed.The backfat thickness growth rate(BGR)characteristic associated with individual variability in pigs was defined.(2)A non-contact measurement model for measuring the BF of pregnant sows,based on image and BGR,was built.A self-organized CNN model was built to extract features from the collected images of pregnant sows,and BGR feature,which considered the heritability of BF,were proposed and combined with SVR regression to build the CNNBGR-SVR-based measurement model for BF.The comparison models based on different features and different regression methods were also built respectively.After tested on ten test sets,the MAE,RMSE,and MAPE of the CNN-BGR-SVR model for measuring BF in pregnant sows were 1.21 mm,1.50 mm,and 7.76%,respectively,and the R~2 value was 0.72,which were all better than the model using different features and regression.This indicated that the use of CNN to extract features from images can effectively replace manually defined features,that BGR feature can help to improve model accuracy,and the non-contact measurement method for measuring pregnant sows BF based on images and BGR was feasible.(3)A hybrid model based on CNN and Vision Transformer(ViT)was built for noncontact measurement of pregnant sows BF.To further improve the accuracy of the noncontact measurement model for BF,a ViT structure with self-attention was introduced based on the CNN model to enhance the global relationship capture capability of the model while retaining the ability of the CNN to extract local features.At the same time,a depthseparable convolution and a lightweighted self-attention were introduced to reduce the large number of parameters and computational effort associated with the ViT structure,making it more suitable for small datasets.The comparison model based on the different structures were also built.After tested on five test sets,the CNN-ViT model had the MAE of 0.84 mm,RMSE of 1.05 mm,MAPE of 4.90%,and an R~2 of 0.74.All indicators outperformed the models based on different structures,and were improved compared to the CNN-BGR-SVR model,with better accuracy and generalization performance.