Re-identification Of Golden Monkey In Nature Scenes

The golden monkey is a national first class protected animal and belongs to endangered species.At this stage,non-invasive selecting methods can be used to obtain a large number of images and videos of golden monkeys in natural scenes.Based on these images or videos,the biometric information of golden monkeys is studied.However,for the task of tracking and re-identification of a particular golden monkey in multiple videos or images,it is difficult to finish the detection and identification tasks using the traditional approach.Based on the theory and methods of deep learning,the method of detecting and re-identification of the golden monkeys in natural scenes is presented.The region of the golden monkey is detected in the original natural scene images taken by multiple cameras,and some obtained monkey patches are used as the query example,then these golden monkeys are retrieved from the image gallery.In this way,the golden monkey re-identification in different scenes can be completed.This thesis firstly expounds the significance and research status of the re-identification of golden monkeys in natural scenes,and summarizes the related theory of re-identification work,Generative Adversarial Network(GAN)and Fully Convolutional Network(FCN).Then the FCN is improved to detect the body of the golden monkey in the natural scenes,and a re-identification algorithm based on Cycle GAN data enhancement is propose for the golden monkey.The body segmentation lays the foundation for the re-identification of the golden monkey,and the data enhancement based on the Cycle GAN is aimed at the problem of data shortage,and effectively improves the accuracy of the re-identification of the golden monkey.Finally,the functional software for the re-identification of the golden monkey in the natural scenes based on Qt is designed and developed.(1)For the problem of the original FCN network segmentation results in which the body of the golden monkey is incomplete,the loss function DWL(Distance-Weight Loss)is designed to optimize the FCN network model.The loss function focuses on the selected golden monkey interest region in the whole image,and combines the distance information from the center to the edge of the ROI to set corresponding weights for different pixels.Therefore,the positional constraint information of the golden monkey truncus to hair is combined,which effectively improves the accuracy and integrity of the golden monkey segmentation result in the natural scene images,and finally can generate the golden monkey detection frame that meets the requirements of the re-identification.(2)For the problem of data shortage in the golden monkey re-identification task,the Cycle GAN network is used to generate winter style data to achieve data enhancement.Since the generated data will have a negative impact on the accuracy,the loss function OGVL(Original-Generated Vary Loss)is designed to distinguish the generated data between the real data,so that the network model pays more attention to the real data during the training process.Experiments show that the data enhancement method can effectively improve the accuracy of golden monkey re-identification,and the OGVL function has the effect of reducing the impact of noise from the generated data.(3)According to the research of golden monkey detection and re-identification algorithms,the re-identification software of golden monkey in natural scenes based on Qt platform is designed and developed.The software consists of two main functional modules,which first complete the detection of golden monkey in natural scene images,and then re-identify the obtained patches,so as to realize the function of re-identification the specific golden monkey individual in the original natural scene image.