【摘要】 及时检测火焰和烟雾，对火灾预警具有重要意义.实时检测Transformer(real-time detection Transformer, RT-DETR)作为一种前沿的模型架构，可以进行实时目标检测.针对该模型参数量大、计算复杂的问题，提出一种基于改进RT-DETR(improved real-time detection Transformer, iRT-DETR)的轻量化火灾检测方法.采用StarNet替换RT-DETR的主干网络以减少模型参数量，并在主干网络中添加注意力机制以提升关键特征的利用效率；提出基于卷积神经网络跨尺度特征融合模块(CNN-based cross-scale feature-fusion module, CCFM)以减少模型计算量，并提高检测的精度；通过RMTRepC3卷积块以提高对小目标火焰检测的能力.改进模型iRT-DETR在自建数据集和公开数据集上分别进行了测试，实验结果表明，改进后的模型在两种数据集上的参数量均减少了51.8%,计算量均降低了59.9%,平均精度均值(mean average precision, mAP)分别提升了1.6%和1.0%,有效地提高了火焰和烟雾检测性能.更多还原

【Abstract】 Timely detection of flame and smoke is of great significance for fire warning. As a cutting-edge model architecture, RT-DETR(real-time detection transformer) can perform real-time target detection. Aiming at the problem of large parameters and complex calculation of the model, a lightweight fire detection based on iRT-DETR(improved RT-DETR) is proposed. Firstly, StarNet is used to replace the backbone network of RT-DETR to reduce the number of model parameters, and attention mechanism is added to the backbone network to improve the utilization efficiency of key features. Secondly, an improved CNN-based cross-scale feature-fusion module(CCFM) is proposed to reduce the computational complexity of the model and improve the detection accuracy. Finally, RMTRepC3 convolution block is proposed to improve the ability of small target flame detection. The improved model iRT-DETR is tested on the self-built data set and the public data set respectively. The experimental results show that the parameters of the improved model on both data sets are reduced by 51.8%, the calculation amount is reduced by 59.9%, and the mean average precision(mAP) is increased by 1.6% and 1.0% respectively, which effectively improves the performance of flame and smoke detection.更多还原