A hierarchical Transformer network for smoke video recognition

During fire incidents, the quick and accurate identification of smoke is crucial for issuing early warnings and reducing the risk of fire. This paper proposes an accurate efficient smoke video recognition network based on a novel hierarchical Transformer architecture. We design the SoftPool-based multi-head self-attention (SMHSA) module, which performs self-attention operations on shortened sequences. This approach facilitates the extraction of global features across various smoke patterns while reducing computational complexity and preserving essential feature information. Our hierarchical network architecture integrates SMHSA modules progressively, enhancing the modeling of global dependencies among image patches of different scales. Specifically, shallower layers are dedicated to analyzing small-scale patches, while deeper layers focus on larger-scale patches. This structure optimizes the model's ability to capture multi-scale information, which is critical for accurate smoke recognition in video sequences. Additionally, the self-attention mechanism is implemented on sequences of progressively decreasing lengths, leading to a significant reduction in computational complexity. To support thorough evaluation and advancement in this field, we have created a dedicated smoke video recognition dataset (SVRD) that includes a wide range of scenarios and smoke patterns. Using the SVRD, we conducted extensive experiments to validate the effectiveness of our approach. Our findings clearly demonstrate that the proposed network achieves superior accuracy in smoke recognition while maintaining significantly lower computational costs compared to existing methodologies.