Image Based Individual Identification Of Endangered Animals

Endangered animal conservation is critical for maintaining biodiversity.If endangered animals are not effectively protected,it will lead to imbalanced ecosystems.Endangered animal conservation is becoming more and more difficult due to global warming and habitat losing.Individual animal identification can efficiently help relevant researchers know the behavior and living state of animals and better protect them.Individual animal identification methods include artificial visual inspection,collecting animal biological samples,attaching sensors to animals,and implanting microelectronic chips into animals.These methods suffer from the problems of requiring manual comparison,the malfunction of sensors,and harming to animal.In recent years,researchers began to apply computer vision and image processing to solve individual animal identification with the development of computer technology.But these methods are mainly based on the key body part of specific animals,which require some special conditions to get the local images of key body and are difficult to be extended to other animals.In addition,animals do not cooperate with cameras for image acquisition,which makes the methods based on local images fail.Lastly but not least,endangered animals are rarer than ordinary animals,and the available images of endangered animals are relatively limited,which further increases the challenge of building individual animal identification models.In view of the above problems and challenges,this thesis takes two representatively endangered animals,i.e.,Amur tiger and red panda,as example to construct a practical individual animal identification.The main works and innovations are as follows:(1)This thesis collects and establishes an animal database with 13348 images,which includes bird,cat,cow,deer,dog,horse,leopard,monkey,giant panda,red panda,tiger,sheep and so on.Every image is annotated with the bounding box and category.(2)This thesis introduces the collection and annotation process of red panda images in detail,which provides a reference for the image collection of other endangered animals.After comparison and screening,this thesis collects and establishes a red panda individual identification database,which includes 3941 images of 43 red panda individuals.(3)For the individual identification of Amur tiger,this thesis simplifies the pose of Amur tiger into two categories of heading left and heading right according to the characteristics of Amur tiger.The pose classification can supervise the learning of convolutional neural network.And Amur tiger pose-guided complementary feature learning(Tiger-PGCFL)method is proposed to identify Amur tiger individuals.The experimental results on ATRW database show that the method is better than the state-of-the-art methods.(4)For the individual identification of red panda,this thesis locates the face and tail regions of red panda that are highly discriminative.A special data augmentation method is designed to make up for the lack of data.Meanwhile,the pose of red panda is simplified as whether the face and tail is visible to supervise the learning of convolutional neural network.Finally,red panda pose-guided complementary feature learning(Red Panda-PGCFL)method is proposed to identify red panda individuals.The experimental results on Red Panda database show that the method is better than the state-of-the-art methods.