Wildlife Detection Based On Improved YOLOv5m Algorithm

Wildlife is an essential part of the ecosystem and human society and has significant scientific and ecological values.Therefore,strengthening the protection of wildlife is necessary to maintain the ecological balance and promote the sustainable development of human society.However,the traditional methods of artificial wildlife monitoring require a lot of workforce and material resources.Therefore,people began to look for new methods of wildlife monitoring.With the emergence of infrared cameras and the continuous development of target detection technology,the use of infrared cameras for picture acquisition and target detection technology to monitor and identify wildlife has become the mainstream method in the field of wildlife protection,gradually replacing traditional monitoring means.This method dramatically reduces the consumption of human material and financial resources and is significant to wildlife conservation.In addition,with the development of the target detection field,the YOLOv5 m detection algorithm emerged,recognized by more and more scholars for its efficient and accurate detection characteristics,and is widely used in wildlife monitoring.However,most of the research has some limitations,such as the detection accuracy of the wildlife detection model being low,the dataset animal categories being relatively single,detection scenes being limited,and other problems.In addition,due to the changeable scenes and complex lighting of wildlife images,the YOLOv5 m algorithm has some degree of missing and wrong detection behaviors,so improving the detection accuracy of YOLOv5 m for complex and changeable wildlife images has become an urgent problem to be solved.To solve the above problems,this paper constructs a wildlife dataset with rich detection targets and great detection scenarios and improves the YOLOv5 m algorithm.In this paper,we propose the Att-YOLOv5 m detection algorithm based on the attention mechanism and the Inv-YOLOv5 m detection algorithm based on the involution operator.Each performance index of the proposed algorithm is improved,and the main contributions of this paper are as follows:(1)Constructing a wildlife dataset of the Wuling Mountains in Yunnan.In this paper,the wildlife images captured by infrared cameras are pre-processed.The wildlife dataset containing six types of animals and 6041 images is finally obtained by filtering and classifying the original images and labeling them.(2)The attention mechanism-based YOLOv5 m detection algorithm Att-YOLOv5 m is proposed.Firstly,we replace the Focus module in the feature extraction module of YOLOv5 m model to improve the feature extraction ability of the detection network for the input image;secondly,the feature fusion part is improved by embedding the self-attention mechanism into the feature fusion module to improve the network’s ability to focus on important location pixel information.The improved Att-YOLOv5 m improves each index compared with the original YOLOv5 m algorithm,with an overall Recall improvement of10.8%,m AP@0.5 improvement of 5.1%,and overall accuracy improvement of 2.8%.The improved network has a much higher detection capability for targets.(3)Inv-YOLOv5 m detection algorithm based on the involution operator is proposed.Based on the Att-YOLOv5 m algorithm,an involution operator with feature fusion operation is introduced in the feature fusion part to improve the concat splicing operation of the original network.Moreover,the weighting operation is applied to different feature layers according to the importance of the features to give higher weights to the essential feature layers and make the network focus more on the key information.The m AP@0.5 value of the improved Inv-YOLOv5 m algorithm is 94.9%,and the detection results are all the better than those of the Att-YOLOv5 m algorithm.Compared with the original YOLOv5 m algorithm,the m AP@0.5 of Inv-YOLOv5 m is improved by 6.1%,and the accuracy and Recall growth is above 10%.