Research On Non-contact Weight Measurement Of Pre-weaned Piglets Based On RGB-D Images

China is the world’s largest producer and consumer of pork,with the pig industry being an important component of the country’s livestock industry.With the increasing scale of pig farming,the demand for advanced breeding equipment in large-scale farms has also been rising.The integration of information technology with various industries has greatly promoted the development of intelligent pig farming from automation.In large-scale farms,monitoring the growth of piglets in the farrowing house is an important part of pig farming,and the weight information and corresponding growth curve of piglets are important indicators of their health status.Currently,several large-scale breeding companies use RGB-D image-based computer vision systems for weight prediction in piglets during rearing and fattening stages,but there has been no research or application of growth prediction in the farrowing house.Piglet weight measurement currently relies heavily on manual weighing,which not only consumes a lot of manpower but also poses stress risks to the piglets.Additionally,without ear tags at birth,it is difficult to accurately identify piglets and predict their weight before weaning,which leads to difficulty in timely culling of weak and unhealthy piglets that do not meet weight requirements.To address these issues,this study uses deep learning to predict piglet weight using RGB-D images to achieve contactless and intelligent measurement of piglet weight.This allows for the development of AI-based pig growth monitoring technology,which can improve animal welfare,reduce waste and emissions from pig farms,lower labor costs,and accelerate industry-scale development to provide scientific decision-making for production.The study collected daily weight data of 373 piglets in two batches,along with 69,723 corresponding RGB and depth images,to build a dataset.An instance segmentation model was constructed using the captured piglet RGB images as the raw data,and a hand-written digit recognition model was built based on manual annotations of digits on the piglets’ backs.A piglet weight prediction model was developed using the depth images to achieve intelligent weight measurement of piglets.The main research results are as follows:(1)Using piglet RGB images as the original dataset for individual piglet segmentation models,the model achieved a validation set m AP of 78.36% and Io U of 81.77%,and a test set Io U of 82.75%.After segmentation,a total of 154,720 individual piglet RGB images were obtained from 69,723 original RGB images.(2)A hand-written digit recognition model based on piglet RGB images was constructed using manually labeled numbers on the back of each piglet.The model achieved training set accuracy of 99.89%,validation set accuracy of 98.3%,and test set accuracy of 97.95%.(3)Using the segmentation and hand-written digit recognition models,a total of87,276 RGB-D images containing accurate weight information were segmented and recognized with a confidence threshold of 99%.(4)A weight prediction model based on piglet depth images was constructed,The model achieving an average absolute error of 0.339 kg,mean squared error of 0.228,coefficient of determination of 0.899,and correlation coefficient of 0.95 on the test set.