TIR: A Two-Stage Insect Recognition Method for Convolutional Neural Network

Recognition of insect images has been a challenge work due to variation in appearance within a category and similarity between classes. Although it can be regarded as a fine-grained vision classification (FGVC) problem, the nature of insect metamorphosis, that insects within the same class may have very different form at different growth stage, makes it diffierent from other FGVC problems. In this paper, we first refine the IP102 dataset and build IP102-YOLO, an adjusted insect dataset which is more suitable for recognition, and propose a Two-stage Insect Recognition method for convolutional neural network (CNN), namely TIR, to improve its performance. TIR extracts deep features from insect images, then divides them into K groups by appearance similarity, and trains K recognition heads for CNN, each for a group of deep features. Our experimental results indicate that: (1) our dataset (IP102-YOLO) has better recognition performance with the same algorithm; (2) TIR outperforms the state-of-the-art insect recognition methods; (3) some of the most commonly used backbone CNN models achieve higher accuracy by following our TIR protocol. We will make our new IP102-YOLO dataset and code publicly available at https://github.com/Fengqiqif77/TIR.