Research On Re-identification Technology And Its Application In Wild Amur Tiger Conservation

The wild Amur tiger is a national first-level key protected animal,and more and more attention has been paid to the protection and related research of the Amur tiger.Among them,individual identification of Amur tiger images collected in the reserve is often an important part of Amur tiger research,but traditional manual observation methods are difficult to meet the heavy identification task.This paper is based on the deep learning method to carry out the relevant research on the re-identification method of the wild Amur tiger in the natural scene.In the monitoring scene of the wild Amur tiger,the problem of a certain individual Amur tiger that has already appeared is detected and identified from the monitoring screen.We call it the reidentification of the wild Amur tiger.The re-identification of the Amur tiger can provide strong support for the subsequent in-depth analysis of the survival state,population distribution and biological characteristics of the Siberian tiger,thereby promoting the development of wild Amur tiger protection.The main research work and innovations of this article are as follows:(1)To re-identify the Amur tiger,first of all,it is necessary to detect the individual Amur tiger from the collected pictures.In view of the complex background of the collected images,the occlusion of Amur tigers,and the different sizes of the number,a more efficient single-stage detection network YOLOv3 was selected,and the original model was used to adjust the parameters of the prior frame by using a clustering algorithm,and streamlined The loss function of the model effectively improves the accuracy and efficiency of Amur tiger detection.This method can better pay attention to the individual Amur tigers in the detected pictures without increasing the depth and width of the network.Experiments on the wild Amur tiger and wild snow leopard datasets show that compared with the original detection algorithm,the detection accuracy is increased by 1.4%and 1.1%,respectively.(2)A method for re-identification of wild Amur tigers is proposed by fusing different frequency domain information of images.This method combines the high and low frequency domain information of the original Amur tiger image and uses the local information of the image for constraint training to construct a novel wild Amur tiger re-identification algorithm.Aiming at the characteristics of the wild Amur tiger image with complex background,strong light changes,and inconspicuous individual details,this paper uses the image high and low frequency domain information to extract complementary features from the original image to improve the fine-grained information of the wild Amur tiger,and then enhance the feature representation ability.At the same time,by using the local branch structure composed of a set of convolution filters in the training phase to locate the local features of the image,and use these local features to perform regular constraints on the global branches,so as to integrate effective local information into the global description and enhance the features.The discriminative ability.Experiments on the Amur tiger dataset and the wild snow leopard dataset show that the accuracy of Rank-1 reached 95.1%and 71.4%,respectively.(3)The local information based on the pose plays a very important role in the re-recognition task.Considering that due to the existence of occlusion in the natural environment,there is a serious lack of posture joint point information in the collected Amur tiger pictures.We propose a method based on part A wild Amur tiger re-identification framework for posture guidance.Using part of the accurate posture joint point information in the training data set,through training multiple losses from the fusion of local and global features,a training strategy for local posture information to guide the learning of the backbone network is constructed.In the training phase,local information can be integrated into global features.To enhance the ability to express the global features of the image,thereby improving the accuracy of the re-recognition task.The accuracy of Rank-1 re-identification can reach 97.0%,which is obviously ahead of the existing wild Amur tiger re-identification algorithm.