Research Of Group-housed Pigs Posture Recognition Based On Deep Learning

Intensification,scale and standardization are the current development trend of China’s pig farming industry,but with the increasing density of feeding,problems have emerged in the development of China’s pig farming industry.The management and epidemic prevention of pig farms are facing great challenges.Disease,plague is an important reason affecting the sustainable development of China’s pig breeding industry,Therefore,how to detect and solve the problem of pig disease in advance is particularly important.pig posture change is usually a precursor to pig disease,research based on deep learning pig posture recognition technology,Abnormalities in pigs can be noticed sooner and factors that endanger the health of pigs can be identified in advance,reducing the incidence of disease in pigs.disease infection rate,promote the development of China’s pig industry,However,the complex environment of pig farms,as well as objective factors such as mutual contact and shielding between individual pigs,greatly increase the difficulty of herd posture recognition.In this paper,herd pig images are used as research objects,image instance segmentation method as well as mechanism of synergistic attention are integrated with convolutional neural networks to construct a herd pig pose accuracy recognition model and improve the precision of herd pig pose recognition,and a herd pig pose recognition system is implemented based on this model.The key studies covered several areas as follows:(1)Construction of pig instance segmentation data set.Aiming at the lack of segmentation open data set of pig instances,HD cameras installed in farms and manual photography were used to obtain images and videos of pigs standing,sitting,lying and kneeling posture.In the pre-processing process,the Key frames from the video footage are extracted,and the captured and extracted images are subjected to manual filtering for indicators that include blurriness and sharpness.Image enhancement techniques are then used to improve the pig image dataset.Finally,Labelme labeling method was used to label individual pigs in each image,and the final segmentation data set of pig instances was obtained.(2)Study the segmentation method of pig case based on deep learning.Aiming at the unsatisfactory segmentation effect caused by pig adhesion,overlap and other complex scenes,this paper takes Cascade Mask R-CNN as the benchmark network,introduces the feature pyramid network(FPN)and Hr Net V2 network,and constructs the Hr Net V2+FPN composite structure.The Cascade Mask R-CNN feature extraction network is improved to realize highprecision segmentation of pigs in complex scenes such as deep separation,high adhesion and debris occlusion,and provide high-quality data set for pig body attitude recognition.The experiment shows that the improved Cascade Mask R-CNN-Hr Net V2 network has a good case segmentation effect.The average accuracy of the model reaches 0.958 when the threshold of IOU is 0.50 and 0.75,respectively.When the threshold of IOU is 0.05,the step size is 0.05.The average accuracy between 0.5 and 0.95 is 6.8～8.8 percentage points higher than that of the comparison model.(3)Study pig pose recognition method based on deep learning.Aiming at the four pose recognition problems of standing,sitting,lying and kneeling,the data set of individual pig recognition was constructed based on the pig image results obtained by case segmentation.Then,Coordinate Attention(CA)module was introduced in Mobile Net V3 to build the CA-Mobile Net V3 lightweight pig body attitude recognition network,enhance the feature learning of key parts of pig body,and realize accurate and rapid recognition of individual pig attitude.The experimental results show that,compared with VGG16,Dense Net121,Res Net50 and other network models,the CA-Mobile Net V3 model has the best recognition accuracy in kneeling posture,lying posture,sitting posture and standing posture,reaching 96.9%,99.5%,98.6%,99.1%,respectively.(4)Develop pig attitude recognition system based on PyQt5.In order to facilitate breeders to analyze pig body posture,this paper a pig pose recognition system is developed based on the improved Cascade Mask R-CNN-Hr Net V2 pig segmentation network and CA-Mobile Net V3 pig pose recognition network.The system has image acquisition function,herd instance segmentation function,live pig individual extraction function,and posture recognition function.