Few-Shot Learning for CNN-based Animal Classification in Camera Traps using an Infrared Camera
Koki Kishi, Masako Kishimoto, Sulfayanti Faharuddin Situju, Hironori Takimoto, Akihiro Kanagawa · 2023
This study evaluated the effectiveness of typicalfew-shot learning methods for training CNN models inanimal classification using a small dataset of imagescaptured in infrared light. Specifically, we evaluated theeffectiveness of transfer learning and data augmentation fortraining of a CNN using only a small number of images.Transfer learning can achieve high accuracy even withlimited training data by utilizing the knowledge transferredfrom a large amount of training data. Here, we focused onthe effectiveness of formula-driven supervised learning(FDSL), which uses fractalgeometry images forpre-training and pre-training. In addition, we investigatedthe use of pseudo-infrared images generated through simplecolor transformation. Similarly, data augmentation canimprove model performance even with a limited number ofimages through image processing that virtually increasesthe number of images by virtually increasing the number ofimages through image processing techniques that alter theshapes and tones of the original images. We verified theeffectiveness of this method by applying typical dataaugmentation methods to the classification of infraredanimal images.