Fire and Smoke Detection using Deep Learning

Module: 7053SCN — Deep Learning and Computer Vision (Coventry University)

This repository compares two deep learning approaches for fire and smoke detection:

A Custom CNN (trained from scratch) for image classification — fire/smoke present vs none.
YOLOv8 for real-time fire and smoke object detection.

The two models are compared on accuracy, precision, recall, F1-score, training time and inference speed, demonstrating both image-classification and object-detection approaches for a real-world fire and smoke detection system.

Dataset

D-Fire dataset (Kaggle: sayedgamal99/smoke-fire-detection-yolo) — over 21,000 images with YOLO-format bounding-box annotations for fire and smoke. The classifier uses image-level labels derived from the boxes; the detector uses the boxes directly.

Notebook

notebooks/fire_smoke_detection.ipynb runs the full pipeline: data preparation, the Custom CNN (training, evaluation, confusion matrix, inference speed), YOLOv8 (training, mAP, detections), and a final comparison table and inference-speed chart.

How to run

Open the notebook in Google Colab and select a GPU runtime.
Add your Kaggle credentials (username and key) in the marked cell.
Run all cells. The dataset downloads automatically.

Reproducibility

Random seeds are fixed and library versions are printed in the setup cell.

Gogulab/7053SCN-Fire-Smoke-Detection

About

Resources

Stars

Watchers

Forks

Releases

Languages