Abstract: Aiming at the problems of traditional fire alarm can not be warned before the occurrence of fire, the effect is not good in extreme weather such as night, and it is limited by complex environment, a fire alarm algorithm based on infrared and visible light images fusion is proposed. A two-branch attention structure is designed and proposed in a generative adversarial network (GAN). One branch extracts more robust feature information through the dense residual network, and the other branch makes up for the lack of spatial information through the efficient coordinate channel attention group (ECCAG) to maximize the acquisition of more high-frequency detail features, and designs and proposes a regulation loss as a loss function, and obtains the fusion image by improving the GAN algorithm. Finally, according to the proposed fire warning algorithm, whether there is a fire hazard is judged. The experimental results show that, the average accuracy of object detection in the fusion dataset obtained by the improved GAN algorithm is 96.19%, which is improved by 11.09?percentage points and 6.2?percentage points compared with the infrared dataset and the original GAN algorithm dataset, respectively, and the accuracy of flame hazard detection on the TNO and LLVIP datasets of the public dataset is 97.45%. The results show that the fire warning algorithm can warn in time when no fire occurs, and can obtain significant detection effects for different scenarios.

Key words: generative adversarial network, image fusion, early fire warning, two-branched structure, attention mechanism

摘要： 针对传统火灾报警在夜间等极端天气下效果不佳，受限于复杂环境等问题，提出一种基于红外与可见光图像融合的火灾预警算法。在生成对抗网络(GAN)中设计并提出双分支注意力结构。其中一条分支通过密集残差子网提取更多鲁棒的特征信息，另一条分支通过注意力子网(efficient coordinate channel attention group,ECCAG)弥补空间信息的缺失，以最大限度获取更多高频细节特征，设计并提出了一种调节损失作为损失函数，通过改进GAN算法得到融合图像，根据提出的火灾预警算法判断是否存在火灾隐患。实验结果表明：改进GAN算法得到的融合数据集目标检测的平均准确率为96.19%，相较于单一红外数据集与原始GAN算法数据集的目标检测平均准确率分别提高了11.09个百分点与6.2个百分点，在公开数据集TNO与LLVIP数据集上测试火灾患检测准确率为97.45%。结果表明，火灾预警算法可以在未发生火灾时及时预警，针对不同场景都可得到显著的检测效果。

关键词: 生成对抗网络, 图像融合, 早期火灾预警, 双分支结构, 注意力机制