.png){kind=icon}
.png){kind=icon}
Return
Masked autoencoder-based vision framework for robust fire detection in complex environments
.png){kind=icon}
PRE
AI
DOI：10.1016/j.psep.2025.108019.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Vision-based fire detection has become an increasingly important focus in computer vision, driven by the growing need for early warning systems and public safety in surveillance environments. While conventional models have primarily relied on color-based features to distinguish fire from background, maintaining high detection accuracy while ensuring computational efficiency remains a persistent challenge, particularly in real-time surveillance systems. To address this, we introduce a novel fire detection framework grounded in masked autoencoding and Vision Transformers (ViT), designed to balance detection performance with scalable deployment. Our architecture leverages self-supervised learning to reconstruct masked visual regions, enhancing the encoder’s ability to capture fine-grained fire cues in complex scenarios. The integration of global attention and hierarchical context modeling enables the system to distinguish between fire and visually similar non-fire patterns, such as reflections and artificial lighting, under diverse environmental conditions. Unlike prior models that are sensitive to background noise or rely heavily on channel saliency, our approach learns robust representations through reconstruction objectives, eliminating the need for hand-crafted modules. Extensive experiments conducted on five benchmark datasets: BWF, DQFF, LSFD, DSFD, FG and DFAN demonstrate consistent improvements over existing methods, with notable gains of 2.5% on BWF, 2.2% on DQFF, 1.42% on LSFD, 1.8% on DSFD, 1.14% on FG and 1.10% on DFAN. The proposed model also maintains computational efficiency and generalizes effectively across a wide range of fire conditions, supporting its deployment in practical, real-time systems.
Journal
IF:
7.8
Papers:
9.5K
Citations:
3.8W
Organization
View more organizations