FlameVideo: spatiotemporal dehazing-attention networks with benchmark datasets for robust abnormal flame detection

Abnormal flame recognition plays a critical role in daily fire safety by enabling early detection of hazards and supporting timely safety alerts. Despite its importance in safety monitoring and fire prevention, accurately identifying abnormal flames remains challenging due to the complex nature of flame dynamics and the scarcity of comprehensive video datasets. Moreover, most methods conduct research based on image data, which lacks temporal-evolution information and thus cannot capture the time-dependent characteristics of flames. To address these challenges, we propose a novel architecture based on the Inflated 3D ConvNet (I3D) framework, which is enhanced with a dehazing module and an attention mechanism. This design enables effective extraction of spatiotemporal features from video sequences, improving recognition performance. Additionally, we introduce two new video datasets—FlameVideo01 and FlameVideo02—constructed from publicly available flame images. These datasets comprise 5,209 and 8,410 video samples, respectively, and are tailored for abnormal flame recognition tasks. Our enhanced I3D model achieves over 92% accuracy on both datasets, demonstrating performance comparable to that of image-based recognition approaches. These results validate the effectiveness of our method, which has strong potential for application in urban, forest, and other fire-prone environments.