Sheep Identification with Distance Balance in Two Stages Deep Learning

Abstract

Sheep identification plays an important role in precision sheep breeding. In this paper, we propose an identification method based on computer vision and deep learning. The whole procedure is divided into two conjunction stages. Sheep face detection stage searches the candidate area in the image firstly. We adopt YOLO framework and optimize it with an overlap-free model. A comparative study is analyzed on various backbone networkse.g., EfficentNet, ResNet, etc. The second stage extracts the feature vector of the candidate sheep face area into a 2048 vector. The distance balance policy is used to maximize the distance of different sheep faces and to minimize that of the same ones. We also introduce a dataset of 547 sheep containing over 5000 images that are suitable for sheep identification task. All images are annotated with the bounding box to mark the position and size of the sheep face, and tag with the ID number as the identity of the sheep. Experiments on this dataset show that our proposed method can achieve the sheep identification accuracy in 85%, and the EfficientNet as the backbone can overwhelm the other networks. For an original image, it can detect the sheep face in 0.024s, and get the recognition result in 0.16s. Compared with the traditional RFID and ear-tag technology, the fast speed and high accuracy of sheep face identification show the great potential of using it in husbandry in the near future.