Real-Time Fire Detection Algorithm Based on Improved RT-DETR

Fire detection is an important means to safeguard public safety and reduce property loss. Traditional fire detection methods are susceptible to environmental interference and difficult to achieve early warning in open scenarios, and Existing deep learning-based methods struggle to balance detection accuracy and real-time performance effectively, and have poor adaptability to the diversity of flame patterns. In order to solve these problems, an efficient real-time fire detection framework ISD-DETR is proposed in this paper. Firstly, the BasicBlock in ResNet18 is improved by using the inverted residual shifted block, which combines the depth-separable convolution with the extended window multi-head self-attention mechanism to enhance the multi-scale feature extraction capability and reduce the computation amount. Secondly, a feature fusion module based on semantic and detail injection is introduced to optimise the multi-scale feature interaction through the spatial-channel attention mechanism, which significantly improves the feature expression capability. Finally, a diverse branching block is proposed to enhance the feature modelling capability in the training phase through the multi-branching structure, while maintaining the inference efficiency by using structural reparameterisation. Experimental results on the fire image dataset constructed in this paper show that the improved model achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 1. 6 \%, 5 1. 7 \%, 9 2. 1 \%, 8 7. 1 \%, ~ a n d ~} \mathbf{7 1. 3}$</tex> for mAP50, mAP95, Precision, Recall, and FPS, respectively, and the GFLOPs are reduced by 13.5% compared with the benchmark model, which balances real-time and accuracy.