Abstract

Fires pose a serious threat to life and property, making early flame detection critical for reducing fire losses. However, existing single-band flame detection methods cannot fully exploit complementary spectral information and are prone to false alarms in complex environments. To address this issue, we propose a Dual-Band Flame Attention Network (DBFANet), which consists of a visible-light channel, a near-infrared channel, and a fusion channel. The visible-light and near-infrared channels employ DAB-DETR for flame detection, while the fusion channel adopts a multi-level feature fusion structure with spatial and channel attention mechanisms to enhance effective fusion information. In addition, a Dual-Band Flame Deep Context Fusion Module and a Flame Texture Information Aggregation Module are designed to improve cross-band feature representation and multi-scale flame perception. A Dual-Band Comprehensive Decision Module is further introduced to integrate the detection results from all three channels and suppress false positives under complex illumination conditions. Experimental results on a self-built dual-band flame dataset show that DBFANet achieves average precisions of 95.0% and 93.1% in the visible-light and near-infrared bands, respectively, with false alarm rates as low as 0.013 and 0.025. These results demonstrate the effectiveness and robustness of the proposed method for flame detection in challenging environments.

IPC Classification

Keywords