🌐 Languages: [English] | [한국어]
ToxHerb-Net: Donguibogam-inspired Vision-based Discriminator Network for Poisonous vs Medicinal Herb Classification
(A Donguibogam-inspired image classification model for poisonous vs medicinal herbs)
ToxHerb-Net is an image classification model that distinguishes poisonous herbs (Poisonous) from medicinal herbs (Medicinal) using Donguibogam-inspired poisonous herb identification image data. We selected benchmark models pretrained on large-scale image datasets as backbones, then performed transfer learning on approximately 600GB of poisonous herb identification images to choose the best-performing model. In experiments with an EfficientNet backbone, the model achieved a Top-1 Accuracy of 88.932%, demonstrating its effectiveness for poisonous vs medicinal herb classification.
- Competition: 2021 AI Donguibogam Poisonous Herb Identification Hackathon
- Period: 2021.11 ~ 2021.11
- Host: Ministry of Science and ICT
- Award: 🥉 3rd Prize
- EfficientNet-B4
- EfficientNet-B5
- ResNet-50
- Squeeze-and-Excitation (SE) Block
- PyTorch
- Python
- TensorFlow
- Build and compare models with a focus on EfficientNet-based backbones
- Improve classification performance by applying SE Blocks
- Select the optimal model through transfer learning on a large-scale dataset of approximately 600GB
| Metric | EfficeientNet-b4 with SE blocks |
|---|---|
| Accuracy | 89.932 |
| Loss | 0.426 |
- Explore various freezing and fine-tuning strategies to mitigate overfitting and further improve performance
- Improve generalization by addressing class and subset-level image count imbalance in the large-scale poisonous herb identification dataset, including normalization of sample counts
The code in this repository is released under the Apache License.