Mineral Classification
Image classification model for recognizing 10 types of minerals.
Model
The model is based on facebook/deit-tiny-patch16-224 and was fine-tuned on a subset of the MineralImage5K-98 dataset.
The model classifies the following 10 mineral classes:
- quartz
- topaz
- chalcedony
- cassiterite
- hematite
- agate
- magnetite
- beryl
- silver
- gold
Training
Three experiments were compared:
| Experiment | Learning rate | Test Accuracy | Test F1 |
|---|---|---|---|
| Baseline | 5e-5 | 79.01% | 79.13% |
| Augmentation | 5e-5 | 77.76% | 77.99% |
| Higher learning rate | 1e-4 | 80.54% | 80.59% |
The best-performing configuration used a learning rate of 1e-4 without data augmentation.
Dataset
The original dataset is Nech-C/mineralimage5K-98.
For this experiment, 10 mineral classes were selected. The training set was balanced to 282 images per class, resulting in 2820 training images.
Limitations
The model was trained on a relatively small subset of the original dataset and only covers 10 mineral classes. Performance may therefore be lower on minerals, image conditions, or classes not represented in the training data.