Research On Behavior Recognition Method Of Banna Mini Pig Based On Machine Learning

As a kind of purebred inbred pigs with important biomedical and xenogeneic organ transplantation research value,the breeding model and behavioral significance of Banna Mini Pigs have always been concerned by researchers.However,for a long time,the physiological status monitoring and behavior recognition of Banna Mini Pigs mainly rely on the traditional mode of manual inspection and subjective judgment.In the process of behavior recognition of Banna Mini Pigs,the computer vision recognition technology based on camera equipment is an effective solution to reduce the workload of researchers,improve work efficiency and reduce misjudgment and omission.At the same time,the identification of the four basic behaviors of standing,eating,climbing and lying down of Banna Mini Pigs is also helpful for breeders to pay attention to the differences between individual behavior and group behavior of Banna Mini Pigs.Therefore,this study proposes a RC-YOLO v4 target detection algorithm and behavior recognition system.The main research contents are as follows:(1)Sample equalization of Behavioral data of Banna Mini Pig.In order to solve the problem that the unbalanced distribution of training samples in behavior category data may lead to over-fitting in model training,this study proposes a data amplification method of NGD-DCGAN generation antagonistic network,which can randomly generate Banna Mini Pig climbing behavior pictures and complete the supplement to the data set,on the one hand,the number of samples can be relatively balanced,on the other hand,the problem of over-fitting in model training can be avoided.(2)An improved target detection algorithm based on YOLO v4 is proposed.The target detection algorithm based on machine learning is analyzed.By combining residual network Res Net,channel space attention mechanism and YOLO v4,a Res CHS-Net feature extraction network and a complete RC-YOLO v4 target detection algorithm are proposed.This method not only reduces the network parameters,but also ensures the feature extraction ability of the model for the action of Banna Mini Pigs,and improves the recognition efficiency of the algorithm.(3)The application of Transfer Learning.The prior frame clustering optimization of Banna Mini Pig image is carried out,and the Transfer Learning method based on parameter sharing is adopted to adjust the parameter weight of the network model.On the one hand,it speeds up the convergence speed in the process of model training.On the other hand,it reduces the time needed for model training.(4)Development of Banna Mini Pig behavior recognition system.A visual recognition system is developed by using Py Qt5 tools,and combined with the RCYOLO v4 target detection algorithm proposed in this study,a behavior recognition system suitable for Banna Mini Pigs is designed.It is proved that the system can realize the correct output of behavior recognition results and behavior information of Banna Mini Pigs.The experimental results show that,by comparing with Faster-RCNN,SSD,YOLO v4 algorithm,the RC-YOLO v4 target detection algorithm proposed in this paper effectively balances the detection accuracy and efficiency,the recognition effect is also ideal under the condition of slight occlusion.The results show that the target detection algorithm proposed in this study has good robustness.In the contrast experiment,the m AP reaches 93.97%,and the recall rate is 96.27%,the accuracy rate is 89.86%,and the F1 score is 0.95,which is higher than other algorithms used for comparison,indicating that the target detection algorithm has a certain feasibility.