Research On Object Detection Algorithm For Rare Birds On Plateau

Birds are an important part of nature and act as checks and balances on nature’s ecosystems.There are many species of birds that inhabit China,but there are many endangered bird species,especially in the plateau,where there are many first-and second-class bird species protected by the state.The data of rare birds on the plateau collected in this thesis are the rare bird species inhabiting Tibet,mainly due to the special geographical environment and complex and diverse climate of Tibet,which leads to the greater diversity of bird species living in Tibet.Plateau rare birds are a fine-grained dataset,and the habits,shapes and colors of birds have a high degree of similarity.If bird identification relies only on human prior knowledge,it will increase the cost of labor and time.In recent years,the development of computer vision has gradually matured,and object detection and recognition has become one of the hottest research directions in the field of computer vision,providing corresponding technical support for the monitoring of plateau rare birds to a certain extent,promoting the protection of plateau rare birds,saving labor costs,and contributing to the protection of plateau rare bird resources.In this theis,the object detection and recognition algorithm of rare birds on the plateau based on deep learning is carried out,and the basic framework used is the Yolov3 detection network,and the main research work is as follows:(1)Construction of image dataset and preprocessing of rare birds on the plateau.In order to improve the performance of detection,the data augmentation method is used to preprocess the labeled pictures before training,so that the effective annotated data pictures derive more pictures close to the original data value,and the derived explanatory files and the derived pictures are one-to-one correspondence,and then the original data and the derived data are put into the network for training to improve the accuracy of the network generation detection model.(2)A feature extraction network based on mixed attention mechanism is designed.In this theis,the attention mechanism is fused in the last three effective output layers of the backbone feature extraction network of the detection network,so that the network only pays attention to the position of the object of interest in the image,which can improve the accuracy of the object detection model.(3)A detection model for the mixing attention mechanism under the two-stage fine-tuning method is constructed.Since most object detection methods are supervised learning methods,the algorithm requires a large amount of labeled image data,which not only takes time and manpower to label,but also takes a lot of time when training data.In this thesis,a two-stage fine-tuning method for few-shot object detection is adopted,which can reduce the dependence on a large number of labeled data,and train a model with certain generalization ability by using very little labeled data.This thesis also improves on the two-stage fine-tuning method,in which the last three effective output layers of the detection network feature extraction network are added with the attention mechanism when the base class is trained in the first stage,so as to improve the performance of the network generation model.Finally,the experimental results show that the detection model obtained by data augmentation and attention mechanism method has improved the picture detection effect of rare birds on the plateau in the detection network.In the detection network,the two-stage model fine-tuning method based on attention mechanism is also used to perform well on the dataset of rare birds on the plateau.