FlameFinder is a novel deep metric learning system designed to identify fire flames occluded by dense smoke, using paired thermal and RGB imagery captured by firefighter drones. Traditional RGB cameras often fail when flames are hidden by thick smoke; thermal sensors help, but lack segmentation clarity. FlameFinder bridges that gap by training a metric learning model sensitive to flame signature patterns and thermal contrast arxiv.org.

Technical Innovation

The system uses triplet and center loss functions within a deep feature embedding space, combined with attention mechanisms. By training on paired thermal‑RGB images from smoke‑free and smoky scenes, the model learns discriminative representations of flame structures even when they’re partially obscured by smoke. The result: flame detection accuracy improves by around 4.4–7% over baseline (VGG19, ResNet18) in the FLAME2 and FLAME3 datasets arxiv.org.

Impact & Practical Use

In practice, FlameFinder could equip drones or ground robots used by fire services. It recognizes fire behind smoke—critical in situations like building fires or wildland blazes with poor visibility. Enhanced spatial awareness helps teams make safer, faster decisions.

Analysis

This system exemplifies the power of modality fusion and metric learning for safety-critical applications. Rather than simply thresholding thermal images, the model learns semantics—what a flame “looks like” even if partially occluded. The attention modules further emphasize the most informative parts of the thermal‑RGB pair.

Operationally, FlameFinder can reduce false negatives in complex environments. It may be extended to include multispectral data or depth sensors. Combined with real-time alert systems, it could become a standard tool in smart firefighting kits.