Object Detection And Re-identification Model Of Amur Tiger Based On Deep Learning

The Amur tiger is an endangered species,but it is also one of the most promising large feline species to restore its population.Protecting the Amur tiger not only helps to maintain habitat biodiversity,but also helps to protect the common home of humans and the Amur tiger,promoting the sustainable development of human civilization.Object detection and individual identification of the image of the Amur tiger can help researchers to obtain accurate population numbers in time,so as to track the individual dynamics of Amur tiger and analyze and evaluate the geographical distribution of the population,and ultimately formulate a reasonable protection strategy.With the development and wide application of unmanned aerial vehicle(UAV)and nonintrusive camera trap technology,the traditional individual identification method of the Amur tiger faces the challenges such as high labor cost and low efficiency when processing large amounts of video and image data.Therefore,it is very necessary to study a more scientific and efficient identification method for Amur tiger individuals.In this paper,the theory and method of deep learning are used to explore image-based individual detection and re-identification algorithms of the Amur tiger,and the lightweight anchor-free Amur tiger detection network(AF-TigerNet)model and the individual and part feature guided Amur tiger re-identification network(IPFG-Net)model are proposed.Firstly,this paper proposes a lightweight anchor-free Amur tiger detection network(AFTigerNet),which solves the problem of real-time and accurate detection of Amur tigers in natural scenes on resource-constrained devices.The network is based on the design of the GFL network,using improved CSPNet and CSP-PAN as backbone and neck to enhance the feature extraction capabilities,significantly reducing the amount of computation and model parameters;The introduction of random Mosaic and Mixup data augmentation methods improves the robustness of the model;The improved Sim OTA label assignment strategy is used to improve the training stability of the model,and the H-swish activation function is used to enhance the nonlinear fitting ability of the model.The experimental results show that AF-TigerNet achieves55.5% of m AP(0.5:0.95)on the ATRW Amur tiger detection dataset with only 0.617 M parameters and 0.41 B floating point of operations,and achieves a speed of about 74 frames/s on ARM devices,indicating that the AF-TigerNet constructed in this paper is an accurate,fast,and practical Amur tiger detection model.Secondly,this paper proposes an individual and part feature guided Amur tiger reidentification network(IPFG-Net),which solves the problem of low recognition accuracy of existing Amur tiger re-identification methods.The network adds non-local modules to the reidentification baseline network ResNet50,increasing the ability to fit global features;The adoption of strong data augmentation enhances the generalization ability of the model;The use of horizontal flipping to generate false labels avoids mutual matching between left and right sides;A part feature guided module based on maximum mutual information is proposed to fuse local information,and an individual feature guided module is proposed to fuse similar individual features,improving the robustness of features.The experimental results show that the m AP of IPFG-Net achieves 94.4% for single-camera recognition and 81.0% for crosscamera recognition on the ATRW Amur tiger re-identification dataset,indicating that the IPFGNet constructed in this paper is an accurate and effective Amur tiger re-identification model.