Individual Facial Recognition Of Dairy Goats Based On YOLO Object Detection Model

In the process of modern intelligent breeding of dairy goats,fast and accurate identification of individual dairy goats is essential for daily management such as feeding,selection and health.With the popularization of camera equipment and the development of artificial intelligence technology,it is possible to identify dairy goat individuals without stress and contact based on image features.The purpose of this paper is to identify facial images of dairy goats in captivity environment by deep learning technology,to provide a reliable dairy goat individual recognition algorithm,to achieve a more convenient and effective breeding method,and to provide practical reference.The main work is as follows:(1)Construction of a dairy goat face image dataset.The data was collected from the Saanen dairy goat breeding farm.Saanen dairy goats at different growth stages in the captive environment were captured in three stages using several devices,covering different lighting,angle,distance and occlusion conditions under different scenes.The individual characteristics and collected images of the dairy goats were analyzed,and a series of preprocessing operations,such as video framing,image filtering and data expansion,were used.The data collection was divided and constructed according to the shooting period,which contains 3844 dairy goat facial images in 31 categories.At the same time,a subset of dairy goat side faces,young goats and adult goats were constructed to better test the performance of the model and facilitate subsequent research.(2)Dairy goat face recognition based on transfer learning and SE attention.In this paper,YOLO v5 s with good comprehensive performance of accuracy and speed was selected as the model to study the problems of less research on dairy goat individual recognition and the difficulty of high accuracy and accurate identification.Using the transfer learning method based on model fine-tuning,350 sheep images were collected to form a facial detection dataset for the sheep face detection task,and the obtained weight files were used as pretraining weights for the individual recognition task.At the same time,combine the SE attention mechanism module with the C3 module at the end of the feature extraction network,and the SEC# module is to strengthen the integration of different channel features and expand the receptive field.The results showed that the m AP of the YOLO v5s-TSe model which based on Transfer learning and SE attention is 96.68%,that is 1.48% higher than the original YOLO v5 s.(3)Dairy goat face recognition based on Sim AM attention and CARAFE upsampling.Aiming at the problems of small facial differences and difficult feature extraction in dairy goats,a more effective improved algorithm is proposed,which is YOLO v5s-TSC model.Based on the YOLO v5 s model of transfer learning,the Sim AM attention module is used to enhance the feature extraction ability,giving more weight to the features to enhance the information,and the CARAFE upsampling module is used to improve the network feature fusion ability and enhance the learning ability of the model,better restoration of facial detail features in dairy goats,improve model accuracy.The experimental results showed that the m AP of the YOLO v5s-TSC model on the dairy goat dataset and the side face dataset are97.41% and 94.18%,respectively,which are 2.21% and 2.72% higher than the original YOLO v5 s.The model can better cope with the recognition challenges of dairy goats in different environments.In addition,according to the recognition situation of the sub-dataset,the impact of different model improvement strategies and different labeling methods on the recognition results are discussed to verify the robustness of the model.