A few-shot rare wildlife image classification method based on style migration data augmentation

Abstract

Camera traps are widely used in ecological studies to capture detailed shots because they rarely disturb wildlife. However, camera traps can obtain only a small fraction of the images of rare wildlife living in wild habitats and nature reserves, often just a flash of them. Thus, the classification of rare wildlife images faces the problems of small samples, incomplete subjects, and complex backgrounds. We propose a data augmentation method for classifying these images based on a cycle-consistent adversarial network that uses two discriminators and two generators. Images augmented by manual data are input to one generator, and stylized images are input to the other generator, constrained by cycle consistency loss. The two discriminators determine whether a generated image is true or false, resulting in stylized image data. The method can accurately distinguish between similar-appearing species. A model trained by our method was used to classify six rare wildlife species with a classification accuracy of 92.2% and an F1 score of 93.3%. The deep learning techniques widely used in wildlife image recognition require large-sample datasets for adequate accuracy. Our study better solves the problem facing small-sample datasets, when large datasets of rare wildlife are not available. The proposed method can provide a reference for promoting rare wildlife conservation through digitalization and intelligence.