Enhancing real-time fire detection: an effective multi-attention network and a fire benchmark

Over the past decades, fire has been considered one of the most serious natural disasters because of its devastating nature, rapid spread, and high impact on the ecology, economy, environment, and life preservation. Therefore, early fire detection has immense significance in computer vision. However, existing methods suffer from high false prediction rates and slow inference times, which limit their real-time applicability. To bridge these gaps, this study introduces a multi-attention fire network (MAFire-Net) that integrates a modified ConvNeXtTiny (ConvNeXt-T) architecture with channel attention (CA) and spatial attention (SA) modules. These attention modules are integrated after each block of the ConvNeXt-T architecture where the CA module is responsible for capturing dominant channels within the features, leading to highly emphasized feature maps. The SA module enhances the spatial details, enabling the model to distinguish between fire and non-fire scenarios more accurately. Additionally, fine-tuning strategies are applied to streamline the ConvNeXt-T architecture, resulting in an optimized model tailored for real-world fire detection. Furthermore, a comprehensive large-scale fire dataset is developed that encompasses diverse, imbalanced, and challenging fire/nonfire images (both indoors and outdoors). Extensive experiments were conducted to validate the superior generalization capability of the MAFire-Net compared with several baseline architectures using four benchmarks (Yar, BowFire, FD, and DFAN). The experimental results demonstrated that the proposed MAFire-Net outperforms state-of-the-art (SOTA) techniques, demonstrating higher accuracy and faster inference times, which make it an ideal choice for real-time deployment over edge devices.