Research On Object Detection And Identification Of Dairy Cows Based On Deep Learning

Intelligent detection and automated identification of dairy cow in some scenarios by using computer vision will play an important role for the intelligent breading industry in the future.To meet the need of information and intelligent management in the large-scale and intensive dairy cow farms,the study explored the methods of object detection,somatic precision segmentation and identification of dairy cow in complex visual scenes by utilizing computer vision technology and deep learning.The study will enrich the intelligent equipment of dairy cow breeding in Southern Xinjiang and improve its automatic and intelligent management capabilities,and the main works are as follows:(1)To improve the object detection accuracy of dairy cow image in complex scenes,the YOLO V3 was applied to the experiment.The result shows that the undetected rate of single cow image and herd images was 20.69% and 25.00% respectively,and the average detection rate is efficient,0.24 seconds for each image.The study compared and analyzed the detection performance of YOLO V3 in different reliability thresholds,also compared the performance differences between YOLO Tiny-V2 and YOLO V3 in the detection set,and examined the object detection performance of YOLO V3 in different-scale images.(2)To achieve accurate image segmentation of the “adhered” dairy cow body of the herd images in the visual scene,sample data set was marked by LabelMe labeling tool,and the Mask R-CNN was applied to the instance segmentation.The framework was developed based on deep learning with PyTorch,and the Mask R-CNN network was established.The CowNet,an optimized dairy cow image segmentation model,was obtained after training,and the accuracy index was used to evaluate the segmentation results.The result shows that the segmentation accuracy rate of CowNet model reached 84.77% in the sample data set.(3)To realize fast and reliable identification of dairy cow in complex visual scenes,GrabCut and Inception V3 were applied to cow body profile segmentation and identification.Inception V3 network was established based on deep learning with TensorFlow,and an optimized dairy cow identification model was obtained after training.The result shows that the overall average identification accuracy of the FriesianCattle 2017 public data set reached 87.93%,and that of TarimCattle 2020 data set reached 84.54%.Innovation points of the study: Based on the pre-training model of COCO2014,CowNet,an image segmentation model of dairy cow was obtained by data enrichment,manual annotation and tuning training parameters.Images of dairy cow from different perspectives were obtained by constructing 3D model of dairy cow,texture mapping and projective transformation,which can enrich the methods of image data acquisition,meet the need of deep learning for large-scale data,and provide an image enhancement method for difficult collecting target data set.In the future,integration research of identity recognition system,re-identification research of newly added dairy cow and behavioral analysis of dairy cow,and disease prediction research will be carried out.