Research On Face Recognition Of Euler Sheep Based On Deep Learning

Qinghai Province is located on the Qinghai-Tibet Plateau,with rich ecological resources and unique animal husbandry characteristics,and rich in Euler sheep.In order to effectively utilize ecological resources and promote the development of characteristic animal husbandry,the precision breeding of Euler sheep is imperative.In order to alleviate the problems of traditional Euler sheep recognition methods in the breeding process,this paper combines deep learning technology and target recognition technology,and proposes a two-stage Euler sheep face recognition method,including sheep face detection stage and sheep face recognition stage.In this paper,5797 representative frame images are extracted from the Euler sheep video with the assistance of the video screenshot software written by ourselves,and an Euler sheep image dataset is established.In the Euler sheep dataset,there are a total of11594 original frame images and cropped sheep face images,and there are 547 Euler sheep.The Euler sheep dataset is divided into sheep face detection dataset and sheep face recognition dataset.In the sheep face detection dataset,5217 images are used as training set and 580 images are used as test set.In the sheep face recognition dataset,5210 sheep face images of 490 Euler sheep are used as the training set,and 587 sheep face images of the remaining 57 Euler sheep are used as the test set,and there is no situation where the sheep face images of the same Euler sheep are in both the training set and the test set.In the sheep face detection stage,this paper draws on the idea of the YOLO series of algorithms,and based on the comparison of the performance of four backbone networks,a sheep face detection network guided by DIo U is proposed.This sheep face detection network can effectively detect the position and the size of sheep faces,the average DIo U value of detection is 0.66,and the detection speed is 24 milliseconds per image.In the sheep face recognition stage,this paper establishes an Euler sheep face feature library,and proposes a sheep face recognition method that can recognize sheep face images across different datasets.The method uses the Euclidean distance between the feature vectors of sheep face images of different Euler sheep as the basis for judging whether this different sheep face images represent the same Euler sheep.The method does not require retraining the neural network when a new Euler sheep comes along.The recognition accuracy of this method reached 87.6%,and the time required to recognize each sheep face image is 41 milliseconds.In addition,combined with the characteristics of the sheep face recognition method,this paper proposes a method of visualizing heatmap based on the difference between feature vectors,gives the region of interest of the feature extraction network on sheep face images,and demonstrates the effectiveness of the sheep face recognition method in this paper.