| In recent years,the earth’s ecological environment continues to deteriorate,and the living conditions of birds are getting worse and worse,even on the verge of extinction.All countries in the world are making efforts to protect endangered birds.However,because of the numerous species of birds,judging only by the naked eye is not only inefficient and labor-intensive,but also difficult to be guaranteed with accuracy.In addition,human beings are unable to spend all day in the wild,making it a challenging problem to effectively count the existence and distribution of bird populations.With the development of deep learning,computer vision technology has attracted more and more researchers’ attention.This technology can automatically count the species and quantity distribution of birds in nature,which can save a lot of manpower and material resources compared with manual statistics,and is more conducive to providing important reference information for bird protection and ecological environment change assessment.Therefore,it has important application value to carry out automatic bird identification research.To this end,this paper uses computing technology to carry out research on automatic bird target recognition algorithm by combining semantic segmentation,target detection and target classification.The main work includes:1)In view of the problem that the bird classification data set is not suitable for target detection,and there is no large-scale data set for bird target detection at present,based on the classification data set,this paper proposes a bird target detection data set construction method combining semantic segmentation and enhancement method,and realizes the bird target detection.Firstly,the position of birds in the image is calculated by semantic segmentation.Then the complex real environment is simulated by the random combination method,and the location information is processed synchronously to generate the data set of bird target detection.Finally,the YOLOV3 network is used to realize bird target detection.2)As the YOLOV3 target detection network is unable to effectively deal with the fine-grained classification task among bird subclasses and the classification network is unable to distinguish multiple birds in an image,this paper adds classification network to classify bird subclasses on the basis of 1)to make up for the deficiency of fine-grained image classification.3)To solve the problem that classification models may misjudge other features as target features and cause misclassification,this paper proposes a reliable classification method based on semantic segmentation.First,the background information is removed by segmentation.Then,the training and verification were carried out on the lightweight Darknet-Reference network and the heavyweight Darknet53 network respectively,and the lightweight and heavyweight bird target reliable classification networks were implemented to meet the requirements under different conditions.4)Based on the above work,a bird recognition application system is designed.The system first finds each bird in the image and extracts it,then removes the background,then inputs it to the classification module for classification,and finally returns the detection results.The experimental results show that the bird target detection data set construction method designed in this paper can realize the automatic conversion of the bird classification data set to the target detection data set,and effectively enhance the generalization ability of the model.The YOLOV3 bird target detection model trained by it has a good detection effect.The proposed reliable classification method can further realize the reliable target classification of birds on the basis of the original method,and improve the classification accuracy to a certain extent,and has a good practical application value. |
