Image Classification Method Of Few-shot Rare Wild Animals Based On Data Augmentation And Contrastive Learning

Wildlife is an important part of the ecosystem on which human beings depend,and if we want to do a good job in protecting the earth’s ecological diversity,we must protect wild animals,especially rare wild animals.At present,wild protection cameras have been widely used in ecological research,which can capture more detailed shots and rarely disturb wildlife,but for rare wild animals living in wild habitats and nature reserves,the images obtained by wild protection cameras can still only capture rare wild animals in a flash.Therefore,the classification of rare wild animal images faces the problems of small samples,incomplete identified subjects,and complex backgrounds.In this thesis,by deploying wildlife cameras,we acquired video images of rare wildlife and carried out a research on a small sample rare wildlife image classification method based on data enhancement and comparative learning,which mainly includes the following three aspects of work.First,a small sample of rare wildlife image dataset was constructed.After deploying wildlife protection cameras in a provincial nature reserve in Yunnan Province to achieve real-time monitoring,video images of rare wild animals were found through background recording,and a small sample of rare wild animals dataset was produced by intercepting,cropping and assigning labels,which could be used for small sample image classification research.Second,a rare wildlife image classification method based on data enhancement of cycle generative adversarial network is proposed.By first artificially enhancing the original rare wildlife image,and then using the cycle generation adversarial network for style transfer enhancement,the image enhanced by artificial data is input into one generator,the style image is input to another generator,and the cycle consistency loss is used to constrain it,and the two discriminators are used to judge whether the generated image is true or false,so as to obtain the stylized rare wildlife image data,and then the residual network that has been enhanced twice after image training is classified.The model trained by this method was used to classify six rare wild animals,with a classification accuracy of 92.2% and an F1 score of 93.3%.Third,a classification method for rare wildlife images based on contrastive learning is proposed.By first augmenting the original picture with artificial data,and then using the visual representation contrast learning framework to train the rare wildlife image data enhanced by artificial data,a feature extractor is obtained,and the feature extractor is used to extract the features of the test set image,and then use the support vector machine to classify it.The model trained by this method was used to classify six rare wild animals,with a classification accuracy of 83.26% and an F1 score of 83.32%.In summary,this paper applies deep learning technology to the study of rare wildlife image classification,which not only improves the accuracy of rare wildlife image classification,but also uses the style transfer technology represented by cycle generation adversarial network to generate stylized and practically difficult to obtain rare wildlife images,which helps the protection of rare wild animals.