Research On Re-Recognition Method Of Amur Tiger Based On Deep Learning

The Amur tiger is mainly distributed in the northeast of China,the Far East of Russia,and the mountainous areas of northern Korea.It is one of the top ten endangered wild animal species in the world,and one of the representative large animals located at the top of the food chain in the ecosystem.It plays an irreplaceable and important role in maintaining a healthy natural ecosystem.With the development of human society,the number of wild Amur tigers has decreased sharply.Protecting Amur tigers is of great significance for protecting biodiversity and sustainable human development.Individual identification of the Amur tiger can help researchers track and understand the behavior and living conditions of the Amur tiger,thereby making protection strategies more effective.Traditional monitoring and identification techniques for wild Amur tigers often require a large amount of professional human input,which is inefficient and has certain limitations.In recent years,with the development of computer vision technology and the gradual maturity of technologies such as drones and camera traps,the collection of wild animal images in natural environments has become easier.At the same time,the application of deep learning technology in animal individual recognition has gradually received attention.In order to achieve higher accuracy in individual identification of Amur tiger,this paper uses deep learning technology to study the methods of target detection and re recognition of Amur tiger based on two-dimensional images.The main research contents and results are as follows:(1)Aiming at the difficulties in target detection in natural scene images of Amur tiger,a target detection model for Amur tiger based on improved YOLOv5 is proposed.Firstly,using data enhancement to expand training samples to improve the generalization of the model;Then,aiming at the interference caused by factors such as chaotic background noise,a self attention mechanism is introduced into the feature extraction network to further enhance the model’s attention to the foreground and improve the model’s ability to extract fine grained feature information;Secondly,a deep separable convolutional network is used instead of a conventional convolutional network to make the model run faster;Finally,the EIOU loss function is used to train the model,which accelerates the loss convergence speed and improves the regression accuracy of the prediction frame.The average accuracy of the improved model on the ATRW target detection dataset reaches 96.5%,and the frame processing speed reaches 75.8 frames/s.(2)Aiming at the difficulties in the re recognition of Amur tiger,a new method for re recognition of Amur tiger based on guided learning and feature erasure is proposed.A dual flow model using Res Net50 as the backbone feature extraction network is designed.The model includes a global flow and a local flow.At the same time,a small number of Amur tiger images in the ATRW dataset are segmented and labeled,and a semantic segmentation model is established.The semantic segmentation model is used to obtain the segmentation masks of all Amur tigers,Then,the dual flow model is trained by combining the original image and the foreground image,enhancing the perception of the global flow to the fine-grained information of the Amur tiger through the guided learning of the local flow.A guided feature erasure strategy is adopted to improve the robustness of the feature representation.During the training process,a progressive learning strategy is introduced to accelerate the model training speed and improve the model recognition accuracy.Experiments were conducted on the ATRW re recognition dataset,with a single camera m AP of 92.4% and a cross camera m AP of 73.1%.Finally,a complete re recognition experiment for the Amur Tiger was conducted,with a single camera m AP of 90.6% and a cross camera m AP of 72.8%.