Deep Learning Based Livestock Behavior Recognition In Modern Farming

China is a large country in livestock production,and the development status of livestock industry is closely related to the national life.In order to achieve precise breeding and deal with abnormal conditions in time,it is important to accurately grasp the activity status of livestock.For a long time in the past,animals were identified by manual observation or by wearing sensors.The former is highly subjective and inefficient,while the latter makes livestock prone to stress.The disadvantages of both approaches are obvious.With the improvement of computer hardware level and the development of deep learning,computer vision technology is widely used in the field of livestock behavior recognition.The traditional livestock recognition algorithm uses the classical model to detect the input image directly,but in the face of complex breeding scenes and numerous categories of behavior,this approach generally has the problem of low detection accuracy and cannot meet the needs of ranch use.This paper firstly introduces the research significance and current research status of livestock recognition,the basic principles of object detection algorithm and image classification algorithm.Second,to address the lack of dataset,a multi-behavior dataset of sheep is produced in this paper.In the dataset,the behavior of sheep is divided into normal physiological activities and destructive behavior.The former includes eating,lying and standing,and the latter includes biting,attacking and escaping.To solve the challenges and difficulties of traditional livestock recognition algorithms,this paper proposes a two-stage livestock behavior recognition algorithm based on deep learning,which contains the detection stage and the classification stage.It can accurately recognize six behaviors of sheep.To improve the efficiency of the algorithm,this paper proposes a detection model.Based on a classical network,the multi-scale feature aggregation,attention mechanism,and separable convolution module are imported to it,which makes the network trade off detection accuracy and model size.In the classification stage,due to the unbalanced distribution of samples in the dataset,the classification accuracy of some small sample categories is low.To address this problem,this paper optimizes the loss function of the classification network so that the classification network can achieve the balanced classification results for all categories.Finally,to verify the effectiveness of the two-stage livestock behavior recognition algorithm proposed in this paper,experiments are conducted on it in the multi-behavior dataset of sheep.Experimental results show that it provides good recognition results compared to the traditional algorithm.It also has a smaller model size,which is convenient for the deployment in embedded devices with limited storage.The recognition effect of the algorithm is demonstrated by visualization.