Re-identification Of Amur Tiger Based On Feature Fusion And Domain Adaptation

In recent years,as a key endangered wild protected animal,how to effectively identify and monitor the Northeast Tiger has become an important topic of concern for relevant researchers.Due to the low density,covert behavior,and complex environment of tiger groups in the natural environment,it is difficult to detect them in traditional detection processes.Therefore,it is of great significance to use re identification technology to conduct protective research on them.This article analyzes and optimizes the research status of deep learning based Northeast Tiger re recognition algorithms.The main research content and work are as follows:(1)Currently,most studies on the re recognition of Northeast tigers only use a single strategy to extract local features.However,due to the complex stripe features of Northeast tigers,this method can easily lead to poor model accuracy.Based on this issue,this paper proposes a Northeast Tiger re recognition model CMM-Net(Combining Multi branch and Multi granularity features)that integrates multi branch and multi granularity features.This model combines the multi granularity detail features of the Northeast Tiger with the global coarse granularity features to compensate for the lack of single branch feature extraction ability;At the same time,in order to further improve the model’s representation learning ability,this paper combines multiple classification tasks with deep metric learning Circle Loss to jointly constrain the features extracted from the model.By reweighting between positive and negative sample pairs,the contribution of positive and negative samples to their respective gradients is controlled,allowing the parameters to expand to varying degrees the backpropagation gradients of intra class and inter class similarity during the training process,Thus obtaining a more discriminative model.The final proposed model achieved high accuracy in both single camera m AP and cross camera m AP environments,proving the effectiveness of the Northeast Tiger re recognition model CMM-Net proposed in this paper.(2)This article proposes a YOLOv5 s network optimization model to improve the detection accuracy of Northeast tigers in the wild environment,in response to the problem of detection errors or missed detections in the YOLOv5 s network.Firstly,in the Neck stage,the AF-FPN structure is used to replace the original FPN structure to improve the accuracy of multi target recognition;At the same time,the loss function is improved.In view of the fact that the original loss function does not consider the length width ratio of the boundary box to the prediction box,this paper uses EIo U Loss as the loss function to accelerate the model convergence and improve the accuracy.The accuracy of the YOLOv5 s optimization network proposed in this article for detecting Northeast tigers has been improved compared to the m AP of the original YOLOv5 s network,proving the effectiveness of the YOLOv5 s optimization method proposed in this article for detecting Northeast tigers in wild environments.(3)In response to the issue of inter domain differences in the recognition model for different datasets,that is,if the dataset obtained through object detection methods in natural environments is directly applied to the CMM-Net recognition model,there may be significant inter domain differences between different datasets,which may lead to a decrease in recognition performance when the recognition model trained on one dataset is applied to other datasets,This article proposes a style transfer re recognition model based on TA Net.By migrating the target domain style to the source domain dataset,images with the target domain style are generated to reduce inter domain differences and solve the problem of domain generalization in the recognition model CMM-Net.Finally,the style transfer model was used to re recognize the Northeast Tiger dataset in the wild environment.Both the m AP in a single camera environment and the m AP in a cross camera environment were improved,proving the effectiveness of the proposed style transfer model in solving the problem of re recognition domain generalization.