| Body temperature and its changes are important indicators reflecting the physiological health and welfare status of pigs.The traditional rectal temperature measurement method is time-consuming and laborious,and it is very easy to cause stress reactions in pigs.As a non-contact rapid temperature detection technology,infrared thermography can achieve health monitoring and disease prevention for pigs in intensive production.In this paper,infrared thermography was used to collect temperature data from multiple parts of the pig’s body surface,and then machine learning and statistical analysis methods were introduced to study the dynamic correlation between the pig’s body surface temperature and rectal temperature.A lightweight automatic detection and temperature extraction model for key temperature measuring parts of the pig’s body surface was constructed using computer vision technology.At the same time,a temperature measurement method for fusion on infrared thermograph and visible light images of pigs was proposed.The research mainly included:(1)Under the influence factors of ambient temperature and humidity on infrared thermography,correlations between skin temperature and rectal temperature in six different body surface parts of pigs were analyzed.The forehead and ear root had the strongest correlation with rectal temperature and were determined as the key temperature measurement parts.Then,the rectal temperature prediction model of pigs was constructed based on XGBoost model,which could achieve more accurate prediction of rectal temperature,and its R~2,MAE and MSE on the test set were 0.8982,0.0747 and 0.0208,respectively.(2)Based on the registered visible image data set,a model of detection and temperature extraction of key temperature measurement parts of pig body surface based on Yolov5s-Bi FPN was constructed.The average detection accuracy,model size and detection speed of the model on the test data set were 96.36%,20MB and 100frames per second,respectively.It had the advantages of high accuracy,small number of parameters,small size and fast detection speed.The maximum temperature of forehead and ear root extracted automatically was basically the same as that of manual extraction,with R~2 of 0.988 and 0.974 respectively.(3)A lightweight pig image segmentation model MC-Deeplabv3+was constructed.The mean intersection over union and mean pixel accuracy of the model on the test set were 97.56%and98.76%,respectively,and the model size and parameters were 11MB and 2770196,respectively.It was much lower than Deeplabv3+,Hrnet,and Unet models,which improved the accuracy of pig body segmentation while ensuring light weight,and was more conducive to model transplantation and deployment.(4)The alignment of infrared thermal images with visible images was achieved,and a targeted temperature measurement method for fusing infrared thermal images of pigs with visible images was proposed.The method segmented individual pigs in visible images with an improved semantic segmentation model MC-Deeplabv3+,and then applied the TIF algorithm to fuse them with infrared images.The fused image had excellent performance in both accuracy and visual effect of pig body contour segmentation,realizing the combination of two advantages of pig infrared temperature information and visible light detail information.Meanwhile,the Yolov5s-Bi FPN temperature measurement model based on fused images could achieve better temperature measurement results for pigs compared to visible light images.In this paper,a light-weight automatic detection and temperature extraction method for key thermometric sites in pigs was constructed based on infrared thermal imaging and deep learning technology,meanwhile,a temperature measurement method for fusion of infrared and visible light images of pigs was proposed,which might provide a reference scheme and technical support for the application of non-contact thermometric edge equipment in intensive production of livestock. |
PDF Full Text Request