A light-weight and accurate pig detection method based on complex scenes

With the wide application and rapid development of digital media technology, the interaction between people and computers has become an important part of people's daily life. Pig detection using computer vision is an important technology for realizing fine pig management, real-time monitoring of pig growth and prediction of pig production. In the actual breeding environment, the accurate detection of pigs is difficult, and factors such as target occlusion and small targets seriously affect the accuracy of pig detection. We take a group of healthy pigs in a real breeding environment as the research object and propose a lightweight pig detection method based on YOLOv3-tiny. The method first uses Removal Net to replace YOLOv3-tiny's backbone network, which improves the accuracy and speed of the detection method. Moreover, a new prediction branch is added to the prediction network to improve the detection accuracy for small objects. Then the soft non-maximum suppression(Soft-NMS) algorithm replaces the NMS algorithm in YOLOv3-tiny, which improves the detection ability for occluded objects. Finally, the feasibility and superiority of this method are proved by several groups of comparative tests. The experimental results indicate that our proposed pig-based detection method based on computer vision can provide an effective reference for refined management and real-time monitoring of pigs.