This repository provides the supplementary data for the paper: Insect pest image recognition: A few-shot machine learning approach including maturity stages classification.
Matching networks: is referred to oscarknagg's code.
Prototypical networks: is referred to cnielly's code.
The file Bregman_divergences provides the script for the divergences used as similarity functions. To use it, add the similarity functions to the chosen Few-shot Learning algorithm (Matching and Prototypical networks or another metric-based model).
- Mahalanobis distance;
- Kullback–Leibler divergence (KL-divergence);
- Itakura–Saito divergence (IS-divergence).
The IP-FSL data set, based on curated images from IP102, is composed 97 classes of adult insect pest images, and 45 classes of early stages, totaling 6,817 images. IP-FSL can be downloaded here: https://drive.google.com/file/d/12iguabGCTC2aVpkP8zTDhOeP6HgndeQ5/view?usp=sharing.