Research On Bird Identification Method In Northern Wetland Based On Deep Learning

Animal detection and recognition is one of the main research directions of computer vision and deep learning.It is widely used in animal management departments and ecological protection departments.This article takes northern wetland birds as the research object,focusing on the processing of data sets and convolutional nerves.The improvement of the network,the northern wetland bird network algorithm YOLOv3-Bird is proposed to improve the YOLOv3 network model.In order to enhance the accuracy and capacity of the data set,this paper uses the data set expansion method,uses the unbalanced data set SVM classification algorithm for data classification,uses the SMOTE algorithm to expand the data set and successfully expands the data set to the needs of network training Enhance the research of northern wetland bird target recognition in complex scenes,improve the positioning accuracy and accuracy of northern wetland bird detection in complex scenes,introduce uncertain regression of prediction frames,and guide network learning during network training Predict a more accurate prediction frame,thereby reducing the positioning error of the YOLOv3 algorithm.The MSRCR algorithm is introduced to enhance the graphics degree of image processing,optimize the loss function,and avoid the network level from affecting the update of target characteristics,and improve the structure and design of the network layer The new Darknet-Bird is optimized and reduced the original feature extractor part,making the algorithm more in line with the actual needs of the northern wetland bird data set.The final experimental results show that the improved YOLOv3-Bird algorithm has a recognition accuracy of 86.6% for the experimental target,which is 7% higher than the original YOLOv3 algorithm.The frame number of the YOLOv3-Bird algorithm is37,which is 7 points higher than the original YOLOv3 algorithm frame number,and the detection time is also increased from the original 0.042 s to 0.023 s.In summary,the algorithm can effectively and accurately detect and identify targets,which meets the needs of experiments.