Abstract: Animal recognition with Object Detection (OD) method by using deep learning techniques has gained popularities in biodiversity preservation and “smart farm” program in recent years. Current research mainly focuses on model training and parameter optimization of different forms of livestock in a single scene. To meet the developmental needs of farms in modern agriculture, this paper proposes an image recognition method which preprocesses the morphological characteristics of livestock in different captivity scenarios, by combining Kennard Stone algorithm, K-means II algorithm and deep learning models for different scenes and live pigs of different behavior characteristics. Our results show that the F1-socre, mAP0.5, and mAP0.5-0.95 of the model were 98.48%, 99.27% and 73.03%, respectively. Our research shows significant improvements for promoting smart animal husbandry and saving labor costs for breeding enterprises, which can also speed up research of high accuracy livestock auto-weighing systems and subsequent applied in animal husbandry.

Key words: Deep learning(CNN), YOLOv5, livestock monitoring system, object detection