Research / Insect pest recognition
Class-Balanced Margin Classifiers for Long-Tail Fine-Grained Insect Pest Recognition on IP102
A computer-vision model that identifies 102 kinds of crop-damaging insects from photos, designed to do better on rare species that have few training pictures.
Status: under review at the 29th International Conference on Computer and Information Technology (ICCIT 2026). The results below are reported by the manuscript and have not yet been peer reviewed.
The Approach
The paper proposes CB-Margin, a classifier head that uses cosine similarity with a per-class margin scaled by the inverse effective number of samples. This widens the separation for rare (tail) classes. It is applied to EVA-02 Base (87M parameters) and Large (305M) vision transformers fine-tuned at 448-pixel resolution, combined with multi-scale test-time augmentation.
Dataset
IP102: 75,222 images of 102 species, with an imbalance of about 82 to 1 between the most and least common. The standard split has 45,095 training, 7,508 validation and 22,619 test images.
Results on the Standard IP102 Test Split
- 79.94% top-1 accuracy and 73.16% macro-F1, reported as new state-of-the-art for both.
- +6.82 points tail-class F1 over the linear baseline for the full framework (+2.25 points at matched backbone scale).
Reported state-of-the-art, under review. The paper and its figures are not hosted on this site.