TriFireNet: A multi-stage framework for robust real-time fire and smoke detection

This study presents TriFireNet, a modular framework designed to address the challenges of smoke and fire detection in complex and dynamic environments. TriFireNet efficiently captures multiscale variability, overlapping regions, and irregular patterns by combining bidirectional deep cross-fusion, directed multiscale extraction, and progressive stochastic augmentation. The proposed framework outperforms state-of-the-art models, providing an F1-score of 67.5%, mAP50 of 65.7%, and recall of 64.5%, ensuring minimal missed detections—a crucial feature for safety-critical applications. Precision and computational cost are balanced by TriFireNet’s effective architecture, which offers accurate detection with less computation overhead and a 9.8 ms inference time. These advantages make it the ideal option for safety monitoring and real-time fire prevention, particularly in situations where effectiveness and reliability are critical. • Progressive stochastic augmentation replicates environmental complexities, enhancing robustness to varied lighting and occlusion conditions. • Directional multi-scale extraction employs multi-scale residuals and flow-directed convolutions to capture fine-grained and large-scale fire and smoke patterns. • Bidirectional deep cross fusion framework combines dual spatial pyramid pooling features with bidirectional feature pyramid network to enhance semantic representation and improve detection accuracy. • TriFireNet delivers high recall, effectively detecting overlapping smoke plumes and irregular flame edges.