Face Detection And Recognition Of Goats Based On Deep Learning

In the process of large-scale breeding and breeding of dairy goats,it is an urgent problem to monitor individuals in real time and update population information in time.In order to realize big data management of livestock breeds,individual identification technology is the basic work to be completed.It is a new individual recognition method to record the uniqueness of animal individual features through animal images and recognize them by deep learning method.Based on this,this paper takes dairy goat as the research object,establishes the sheep face data set,completes the sheep face detection and the construction of the sheep face recognition model.1.The literature reviews the existing individual identification of livestock,and summarizes the shortcomings of traditional methods.This paper summarizes the development history of machine learning and deep learning,then selects several common target detection models(r-cnn,Yolo,SSD,yolov3)and image recognition methods(VGg convolution neural network,residual neural network,SVM classifier,softmax classifier)for detailed analysis.2.Taking Xinong saneng dairy goat farm of Northwest A & F University as the research base,35 sheep from three different pens were selected,and a total of 3121 pictures were collected,with about 50-100 photos of each sheep.According to the individual classification of sheep,the background is cut and the semantic segmentation is carried out to obtain the whole body image data set of sheep.Then,the sheep face data set is made by selecting sheep face by manual annotation.3.1002 goat face annotation images are divided into training set(900)and test set(102).The training model is used to detect sheep face in images and videos.Compared with the model trained by yolov3’s own classification standard and dataset,it is found that the model with manual annotation of dataset can select the location of sheep’s face more accurately,and it can complete the detection more accurately in the scene of multiple sheep.After training,the accuracy of the model is 96.94%.4.Using 3121 photos,using triplet loss as the loss function model,using vggface neural network and RESNET neural network respectively,using two kinds of embedding parameters for training comparison,it is proved that when using triplet loss as loss function,higher dimension embeddings(2096)combined with vggface16 network will have bettereffect.A total of 814 pictures of 17 sheep were obtained,616 of which were used as training set and 198 as test set.The training set is put into the network structure for training.After many iterations,the accuracy of the model is as high as 91%.