Research On Key Techniques For Non-contact Automatic Scoring Of Dairy Cow Body Condition

In our country’s animal husbandry industry,accurate monitoring of individual dairy cow information is an urgent need in actual production.Body condition score(BCS)is a common tool that reflects the subcutaneous fat reserve level,the nutrition status,and the reproductive performance of dairy cows,and it is also an effective method,which can satisfy the animal welfare and achieve accurate feeding of dairy cows.By monitoring the body condition score of each cow in different periods and optimizing management strategies in time,the farm can effectively reduce the occurrence of metabolic and reproductive disorders in dairy cows to improve production efficiency and increase the animal benefits effectively.While manual scoring methods are time-consuming,labor-intensive,inefficient,and the contact with dairy cows may cause stress response and affect the animal welfare.Therefore,the non-contact body condition scoring method has research value.Most dairy cow body condition scoring based on traditional methods used the method of marking key points.This method was more complicated in extraction process,difficult to analyze and test,the index of depression degree was not good,and the accuracy could not reach the ideal value.The body condition assessment of dairy cows based on deep learning algorithms can avoid manual marking of key points and reduced subjective errors.Therefore,with the current research progress,the demand for body condition assessment of dairy cows based on deep learning algorithms has become increasingly urgent.Aiming at the low efficiency,high complexity,and low level of automation of the current body condition scoring methods,this paper conducted a research on the key technology of non-contact dairy cow body condition automatic scoring based on deep learning algorithms and computer vision technology.In order to achieve accurate automatic scoring of the cow’s body condition,this study used a depth camera to collect the cow’s back image.First,the target detection algorithm was used to locate the cow’s back target,obtained the cow’s back target and position information,and extracted the detected cow’s back target;Secondly,the extracted target images of the back of the cow were identified to determine the identity of the cow;finally,an automatic scoring model of the cow’s body condition was constructed to achieve accurate assessment of the cow’s body condition.The main contents of this paper are as follows:(1)Data collection and processing of the back image of cows.The Real Sense D435 depth camera was used to collect visible light and depth images of the cow’s back.When collecting the data,two professional cow body condition scorers would evaluate the body condition of each cow and recorded the evaluated scores and ear tags one by one.And saved it to prepare for the subsequent automatic scoring experiment of dairy cow body condition.(2)Target detection on the back of cows.Cow back target detection was the key task of realizing the target location of the cow back and extracting the cow back target as the image to be identified.The SSD model was used to detect the target on the visible light image of the cow’s back,and the model extracted the overall target image of the cow’s back in the image as the image to be identified.(3)Individual identification of dairy cows.Individual identification of dairy cows is the basis for realizing accurate automatic scoring of dairy cow body conditions.At present,individual identification of dairy cows based on deep convolutional neural networks had the problem that when new dairy cows were added to the dairy farm,it took too long to repeat the training.Therefore,this paper proposed a framework for individual identification of dairy cows based on deep feature extraction and matching methods.The identification of individual dairy cows based on this framework can avoid repeated training.Firstly,the trained convolutional neural network model was used as the feature extractor;secondly,the feature extractor was used to extract features and stored the features in the template feature library to complete enrollment;finally,determine the identities of the cows.When a new cow joined the herd,the enrollment can be completed quickly,thereby completing the individual identification of the cow.(4)Automatic scoring of cow’s body condition.A convolutional neural network was used to construct an automatic scoring model of dairy cows’ body condition,and two image fusion strategies are combined to evaluate the cow’s body condition.Firstly,under the first strategy based on the fusion of depth images and RGB images,the fusion data was input to the convolutional neural network;secondly,under the second strategy based on the fusion of depth images and gray-scale hierarchical images,the fusion data was input to to the convolutional neural network.Based on the convolutional neural network,two fusion strategies were combined to realize the automatic scoring of the cow’s body condition within 0 error,0.25 error and 0.5 error.This paper was based on computer vision technology,combined with deep learning algorithms to achieve target detection and individual identification on the back of cows.On the basis of detection and identification,a non-contact accurate automatic scoring of the body condition of cows was realized.It provided guidance for solving the application of individual identification of dairy cows and automatic scoring of body condition of dairy cows in actual production.At the same time,it also provided certain ideas for the needs of individual identification of other animals and automatic scoring of body condition.