Object Detection And Pose Recognition Of Suckling Pigs Based On Deep Learning

China is a big agricultural country.The quantity,consumption,and breeding scale of pork rank first in the world for many years in a row.The healthy and stable development of the pig industry is closely related to people’s livelihood.The National Bureau of Statistics(NBS)showed that the number of live pigs and fertile sows kept growing steadily over the past year.In the traditional breeding industry,the monitoring and treatment of abnormal conditions in the breeding process mainly rely on the subjective judgment of the breeder.The manpower monitoring is time-consuming,labor-intensive and inefficient,and the discovery of abnormal conditions is not timely,which may cause serious economic losses.As a result,the non-contact and efficient features of machine vision technology have been revealed,and the application of intelligent video surveillance methods for farming has emerged.Intelligent video surveillance technology can effectively analyze information and output detection results on various occasions,so intelligent video surveillance technology also has good application prospects in the pig farming industry.Problems may occur in every link and period of pig breeding.With the high efficiency of data processing by computer,machine vision method is used to quickly locate abnormal conditions,make decisions and judgments according to abnormal conditions,and send an alarm to the breeder,to improve breeding efficiency and reduce economic losses.Combining machine vision technology with aquaculture,semi-supervision and automation of aquaculture methods has become a development trend.Based on the deep learning method,we trained models for pig target detection,sow image segmentation and sow pose classification with the self-built pig dataset.The research results of this article were as follows:(1)The target detection framework based on regression method is studied.Based on the deep learning method,the improved YOLOV3 target detection algorithm was used for training through the self-constructed data set of piglets target detection.The accuracy rate,recall rate and F1 value of the optimized algorithm were 92.46%,94.68% and 93.59% respectively for the detection of piglets targets in the complex piggery environment,which could effectively detect the piglets targets and lay a foundation for the subsequent analysis of piglets behavior state.(2)In this paper,the piglets training sets are trained under different target detection frameworks.The detection accuracy of Faster-RCNN,SSD and YOLOv3 models are 90.43%,93.64%and 94.25% respectively.The lightweight network structure is used to improve the detection framework of yolov3,which effectively reduces the training parameters of the model and improves the training speed.The trained model only occupies 26.1MB of memory,which can be used to transplant embedded and handheld mobile devices.(3)Image segmentation method based on deep learning is studied.The image data set of lactating sows was established based on the video data in the pigsty environment,and the semantic segmentation network Deeplabv3+ was used for training.The test set images were used to analyze and evaluate the semantic segmentation accuracy of the results of the training model.It can be concluded that the Deeplabv3+ semantic segmentation model used in this paper can better segment the sow objects in the complex environment of the piggery,laying a foundation for sow pose recognition.(4)In this paper,the image classification method based on transfer learning is studied,and the sow image generated by semantic segmentation model is used as the research material for attitude recognition,which overcomes the problem that the complex environment of piggery affects the recognition performance.Based on the method of model transfer in transfer learning,three kinds of poses of the constructed sow image were recognized,and the four network structures were trained respectively.Compared with the training results without migration,the recognition effect after migration was improved,and the performance of the model before and after migration was analyzed and evaluated in many aspects.The accuracy of MobileNetV2 network structure for sows pose recognition after migration reached 90.23%.