Research On The Key Point Location Method Of Multi-objective Pig Body Size Based On Deep Learning

As an essential indicator of pig phenotype evaluation,body size can not only evaluate the growth status of pigs,but also be used for breeding pigs.Accurately locating the key points of body size is the basis for accurately measuring pig body size.At present,most of the key point locating methods of pig body size usually need to go through a series of complex image processing operations to extract the target contour and then locate the body size key point,which is not only cumbersome to operate but is usually only applicable to a single target in a specific scene,and cannot accurately locate the key points of the body size of multiple pigs in complex scenes(such as aggregation,occlusion).To solve the above problems,this paper took the pigs in the daily monitoring video of the pig farm as the research object,used the method of deep learning to directly locate the key points of the pig’s body size,and improved the object detection method YOLOv5 s and the key point location method HRNet.Finally,a robust multi-objective pig body size key point localization method YOLOv5DA-HRST was proposed.The main work of this paper is as follows:(1)This paper constructed three data sets: the pig posture data set,the singletarget pig body size key point data set,and the multi-target pig body size key point data set.The pig posture data set comprised 7220 images,including three postures: standing,prone lying,and side lying.The single-target pig body size key point data set contained1663 images,and each image was annotated with ten body size key points and a standing bounding box.The multi-target pig body size key point data set contained 150 images which each image was annotated with different postures of the pig and ten body size key points were marked on the standing pig.(2)Based on the object detection model YOLOv5 s,this paper proposed a pig posture detection model YOLOv5 DA to recognize three different postures of pigs(standing,prone lying,side lying).Compared with YOLOv5 s,YOLOv5DA has made three main improvements: first,Mosaic9 data enhancement was used to expand the number of small targets in the data set and improve the model’s ability to detect small targets;second,deformable convolution was added to the feature extraction network to improve the modeling ability of the model;finally,adaptive spatial feature fusion was added in the detection head to learn richer semantic information.The experimental results showed that YOLOv5 DA could accurately identify the three postures of pigs standing,prone lying,and side lying with an average precision(AP)of 99.4%,99.1%,and 99.1%.Compared with YOLOv5 s,YOLOv5DA can effectively deal with occlusion while increasing the mean average precision(m AP)by 1.7%,reaching 86.8%.(3)Based on HRNet,this paper proposed a lightweight single-target pig body size key point location model HRST.The model is a hybrid model consisting of a convolutional neural network and a Swin Transformer Block,which removes the fourth stage of parameter redundancy in HRNet and introduces a Swin Transformer Block based on the self-attention mechanism to capture the constraint relationship between key points better.The experimental results showed that the HRST proposed in this paper achieved an average precision of 90.4% and 91.5% on the single-target pig body size key point dataset and the public amur tiger re-identification in the wild,respectively.Compared with HRNet,the HRST proposed in this paper can reduce the model parameters and calculation amount by 72.8% and 41.7% while increasing the m AP by6.8% and 0.8%,respectively.In addition,compared with the current mainstream key point location algorithms(Center Net,HRNet,HRNetv2,Simple Baseline,Tokenpose),the HRST proposed in this paper can achieve the highest detection accuracy with the fewest model parameters.(4)This paper proposed a multi-objective pig body size key point localization model YOLOv5DA-HRST.This method located multiple pigs’ body size key points by combining the trained YOLOv5 DA and HRST.First,the improved YOLOv5 DA was used to detect different postures of pigs,and then HRST was used to locate the body size key points of standing pigs.The experimental results showed that the YOLOv5 DAHRST proposed in this paper achieves an average precision of 81.5% on the multiobjective pig body size key point dataset.(5)In order to facilitate the operation and use of farmers,this paper used Py QT5 to design the multi-objective pig body size key point location method into an automatic detection software.The software can not only detect picture data,video data,and camera data streams,but also automatically count the number of pigs and the key point coordinates of the body size of each standing pig according to the detection results.