Intelligent Fire Recognition for Surveillance Control Using Cascaded Multi-Scale Attention Framework

Fire incidents cause devastating environmental damage and human casualties, necessitating robust automated detection systems. Existing fire recognition methods struggle with visual ambiguities, illumination variations, and computational constraints, while current attention mechanisms lack hierarchical integration for comprehensive feature refinement. We propose a cascaded multi-attention architecture that combines Multi-Scale Strip Attention (MSSA), Optimized Spatial Attention (OSA), and the Convolutional Block Attention Module (CBAM) to enhance fire detection. MSSA employs three-scale orthogonal strip pooling to capture fire patterns across varying spatial extents through horizontal and vertical feature decomposition. OSA employs optimized convolutions rather than conventional kernels, reducing computational cost while maintaining spatial localization accuracy. CBAM applies sequential channel-level and spatial-level attention for comprehensive feature recalibration. Operating on the EfficientNetB7 backbone, the cascaded design progressively refines representations through complementary mechanisms. Systematic ablation studies validate individual module contributions, while Grad-CAM visualizations confirm precise localization of fire regions. Extensive experiments on the FD and BoWFire datasets demonstrate superior performance compared to state-of-the-art approaches.