Research On Amur Tiger Re-Identification Method Based On Deep Learning

Amur tiger is an endangered species in the world,and it is also the most promising large cat to restore its population.Biodiversity is the synthesis of the ecological complex formed by organisms and their environment and various related ecological processes,including animals,plants,microorganisms,their genes and the complex ecosystem formed by them and their living environment.Biodiversity is the basis for human survival and development,and the protection of biodiversity is an important part of ecological civilization construction in China.On the one hand,protecting tigers can maintain the biodiversity of habitats,on the other hand,it also takes care of the common home of humans and tigers,which is conducive to the sustainable development of human civilization.With the massive deployment of non-invasive cameras in the wild and National Park,traditional identification methods of the Amur tiger have difficulties in processing massive video and image data,which can bring heavy manpower burden and low efficiency.Therefore,it is imperative to study more scientific and efficient identification methods of the Amur tiger.In this paper,deep learning theories and methods are used to study the image-based detection and re-identification algorithm of the Amur tiger.Firstly,the significance and current research status of the Amur tiger reknowledge were expounded.Secondly,the relevant theories of Convolutional Neural Network,Transformer network and Target Re-identification were summarized.Thirdly,an algorithm for natural scene target detection of Siberian tiger was proposed based on Efficient Det.Finally,a re-identification algorithm of amur tiger based on Transformer was proposed and improved.The main work and innovations of this paper are as follows:(1)Aiming at the difficulties of Amur Tiger detection in natural scene images,a detection model based on Efficient Det is proposed.Firstly,a variety of image transfer methods are used to expand the training samples to improve the robustness of the model.In order to reduce the problem of low Amur tiger detection accuracy caused by background interference,an attention mechanism is added into the feature extraction stage to improve the target saliency.Aiming at the problem of missing detection caused by the traditional Non-Maximum Suppression and violent screening anchor in special scene detection,the Soft Non-Maximum Suppression is used to improve the detcetion ability of occluding targets.The improved algorithm has a detection accuracy of 95.0% on the ATRW natural scene detection dataset and can process 19 images per second(2)Using body surface texture information to Re-identificition belongs to fine-grained discrimination.Aiming at solving the problem of Convolutional Neural Network’s convolution and pooling operation which can cause some fine-grained features lose,an Amur Tiger re-recognition baseline model based on Transformer was proposed.(3)In view of the spatial information loss of pixels in the patch splitting process of Transfomer,a patch slidding-window-splitting module was designed to make neighboring patches share part of pixel information,which can be helpful to construct a more comprehensive feature representation of the model.Secondly,in order to avoid the matching phenomenon of very similar individuals with different orientations,the Pose Embedding Module was designed,and the left and right orientations of the Amur tigers were set as learnable parameters,which enabled the model to distinguish the posture of Amur tiger.In addition,MS Loss was introduced into the metric learning stage to enhance the learning ability of difficult samples in the model.Finally,ablation experiments were conducted to verify the performance improvement effect of different modules and methods on the baseline model.The experimental results on ATRW re-identification dataset show that the Patch Slidding-window-splitting Module,Pose Embedding Module and MS Loss can improve the performance of the model,and the improved model improves the mean m AP(mm AP)by 4.9% compared with the baseline model in single camera and cross camera cases.