Individual Identification Of Dairy Cows Based On Deep Learning

Accurate and rapid individual identification of dairy cows is an essential part of automated breeding.It also plays an important roles in guaranteeing cows' health,and improving management capacity and production effectiveness of the livestock farm.Traditional artificial ear tag identification has some weaknesses,such as,low efficiency,easy to wear and occupying human resources.Two-dimensional ear tag,stomach tag,radio frequency identification electronics have the disadvantages of high cost,short identification distance,low serviceability ratio,and causing physical damages to the cows.To solve these problems,based on the analysis of the current domestic and foreign researches of cow identification,this paper adopts computer vision and video analysis technology,and combines the diversity of body spots on the cow back to study a deep learning-based individual dairy cows identification method.The method can provide a contactless method to identify cows with high accuracy and low cost.The main work and conclusions of this paper are as follows:(1)A cow's back video capture platform was built and the cow image preprocessing method was studied.Based on the routine milking process and the actual condition of the milking parlour,software and hardware platforms used to collect videos of cow's back and the video processing platform were designed and established.A total of 48 videos of the cow's back were captured.To reduce noises in the captured video frame images,bilateral filtering,Gaussian filtering,median filtering,mean value filtering and box filtering were conducted respectively.The results showed that the bilateral filtering method removed noises well and the noise removal result had the most powerful similarity with the structure of the original image and the highest signal-to-noise ratio.Brightness,contrast and sharpness of the images were strengthened to improve the image quality.Zhang Zhengyou's checkerboard calibration method was used to calibrate the video camera and distortionless cow images were obtained.The sample diversity was augmented by rotating the cow images and brightness expansion.Since the cows' head is small and change greatly with different cows,only the back region was labeled and a sample set of 36790 frames of 89 cows was established.(2)A cow individual recognition model based on improved YOLO v3 was proposed.First,with Darknet-53 as the backbone,and according to the characteristics of cow and the disadvantage of YOLO v3,YOLO v3 model was improved by optimizing the number and the size of appropriate network anchor boxes,modeling location parameters of the predication box,improving loss function and adding 1×1 convolution layer and residual module.In addition,different identification models obtained by training the network with different iteration times and confidence threshold were evaluated to obtain the optimal improved YOLO v3 identification model.To analyze and verify the proposed identification model,cow individual identification models based on Faster R-CNN and YOLO v3 were established for comparison.(3)The individual identification model of dairy cows was tested and analyzed on the test set samples.The dairy cow test sets obtained from the actual livestock farm were input respectively into improved YOLO v3 model,YOLO v3 model and Faster R-CNN model.The experimental results showed that for the improved YOLO v3 model,the accuracy rate is 95.91%,the recalling rate is 95.32%,m AP is 95.16%,and the average frame rate is 32 fps,all of which are higher than those of the other two models.The results indicated that the improved YOLO v3 model had higher identification accuracy and had the ability to adapt to deformation of the cow.The preprocessing method was proved to improve the identification result.The fall-out ratio of video segment recognition was 3.63%,which could met the farm's requirements for real-time identification of the dairy cows.(4)The identification results of three models under the conditions with different body spot colors and illuminations were compared to verify the robustness of the identification model.Samples with pure black and pure white body spots were identified under the illumination conditions of sidewise sunshine and lamplight.The F1 value of the improved YOLO v3 model was 2-5% higher than that of YOLO v3 model and Faster R-CNN model in identifying dairy cows under different conditions.The result indicated that the proposed modified YOLO v3 model had excellent performance and robustness in identifying cows with different spot colors under different illumination conditions,which can provide technological support to real-time identification of the numbers of individual dairy cows in and out the milking parlour of the dairy farm.