Fire detection models based on attention mechanisms and multiscale features

Fire detection is critical in applications such as fire management and building safety, but dispersion and blurring of flame and smoke boundaries can present challenges. Multiple upsampling and downsampling operations can blur the localisation signals, thus reducing accuracy and efficiency. To address this problem, we propose the AMMF(Attention Mechanisms and Multiscale Features) detection model, which integrates an attention mechanism and multi-scale feature fusion to improve accuracy and real-time performance. The model incorporates a dynamic sparse attention mechanism in the backbone network to enhance feature capture and restructures the neck network using CepBlock and MPFusion modules for better feature fusion. MDPIoU loss and Slideloss are then utilised to reduce the bounding box regression error and address the sample imbalance problem respectively. In addition, parameters are shared by merging 3?3 convolutional branches, which optimises the detection head and improves computational efficiency. The experimental results show that AMMF-Detection can significantly improve the detection speed and accuracy on the public dataset.