tobybreckon/fire-detection-cnn

Implements multiple experimentally-optimized CNN architectures (FireNet, InceptionV1/V3/V4-OnFire variants) offering a 3-4× performance improvement over standard models while achieving 0.93-0.96 accuracy on full-frame binary detection. Supports both global fire classification and pixel-level localization via SLIC superpixel segmentation with OpenCV, operating at 12-17 fps on standard hardware. Built on TensorFlow 1.15 and TFLearn with pre-trained weights available, enabling inference on video streams without temporal information.

570 stars. No commits in the last 6 months.