Real-Time Semantic Segmentation of Fire Using Deep Learning

This project is focused on creating a fire detection model to enhance rescue efficiency with improved accuracy and speed, aiming for real-time processing.

Project inspired by: real-time-fire-segmentation-deep-learning

Our goal was to reproduce the experiment results and test the model on random YouTube videos to evaluate accuracy.

Result 1

Result 2 - Random Video from YouTube

Model Architecture

The encoder consists of a DCNN with 16 convolution layers and an ASPP module (Atrous Spatial Pyramid Pooling). The activation function is a combination of both ReLU and HardSwish. The Lion optimizer and Focal Loss are used.

Results

Loss, MPA [Mean Pixel Average], and MIoU [Mean Intersection over Union] recorder after every epoch

How to Use the Code

1. Clone the repository.
2. Download the dataset files "Images for fire segmentation" and "Masks annotation for fire segmentation" from IEEE Dataport and move them to the 'data' folder.
3. Follow the instructions in the main.ipynb file.