ConFirePatch: Context-based fire detection and segmentation via extracting key patches

Fires pose significant threats to human lives and property. However, existing methods face limitations in edge deployment, accuracy, and efficiency. To address these challenges, we propose ConFirePatch, a new context-based framework for fire detection and segmentation in still images via extracting key patches. Specifically, fire-like pixels are first located by color and texture analysis. Then, key patches centered on these pixels are extracted, and a self-similarity descriptor is introduced to compute their contextual correlation with the surroundings. These correlations are fused with corresponding patches to form context key patches, enhancing fire representation. To fully extract discriminative information from context key patches, we design a Spatial-Channel Attention Network (SCA-Net) and integrate it into the framework. The model enhances feature learning via attention mechanisms, achieving a 0.03MB size for edge deployment. Additionally, to further improve system efficiency, we propose block-based detection and patch-based segmentation methods. Applying ConFirePatch only to fire-like regions significantly reduces resource waste. Experiments on a Raspberry Pi 4B with 4GB RAM using the improved BoWFire+ dataset show that ConFirePatch achieves 0.8800 accuracy and 0.8824 F1-score, with per-image detection and segmentation times of 0.46 and 3.25 s, respectively, outperforming comparative methods and showing strong potential for real-world applications.