|
--- |
|
language: "en" |
|
library_name: "keras" |
|
tags: |
|
- image-classification |
|
- fire-detection |
|
license: "mit" |
|
datasets: |
|
- flame |
|
metrics: |
|
- accuracy |
|
- f1 |
|
model_creator: "CPSquad" |
|
course: "1INF52 (PUCP)" |
|
--- |
|
|
|
# Fire Classification Models |
|
|
|
These Keras models were developed by **CPSquad** as part of a Deep Learning project for the **1INF52 course** at **PUCP**. We trained them on the **FLAME dataset**, which provides UAV-based imagery of wildfires. |
|
|
|
- **DenseNet**: `densenet_final.keras` |
|
- **ResNet**: `resnet_final.keras` |
|
- **Xception**: `xception_final.keras` |
|
- **Ensemble**: `ensemble_model.keras` |
|
|
|
## Hyperparameter Tuning |
|
|
|
Using [Keras Tuner](https://keras.io/keras_tuner/), we optimized: |
|
- Dropout rate |
|
- L2 regularization factor |
|
- Number of layers unfrozen |
|
- Learning rate |
|
|
|
These improvements helped boost performance metrics such as **accuracy** and **F1-score**, allowing us to reach SOTA results on FLAME’s fire/no-fire classification task. |
|
|
|
|
