Liu Qi, Chen Hong, Lin Da
School of Mathematical Sciences, Inner Mongolia University, Hohhot, 010021, China.
Sci Rep. 2025 Jul 1;15(1):20364. doi: 10.1038/s41598-025-06721-3.
Fire detection technology is essential for safeguarding public safety and minimizing property damage. Despite advancements in both traditional methodologies and modern deep learning models, challenges such as suboptimal accuracy and elevated false alarm rates persist, particularly in complex environmental scenarios. This paper introduces the Fire Focused Detection Network (FFDNet), a state-of-the-art flame detection framework that seamlessly integrates classical approaches with deep learning strategies. By leveraging an enhanced Real-Time DEtection TRansformer (RT-DETR) model alongside the Vector Quantized Generative Adversarial Network (VQGAN), our methodology not only enhances flame detection sensitivity and precision but also significantly reduces false alarm frequencies. Specifically, we have integrated a novel loss function, the Innovative Minimum Perimeter Distance IoU (InnMPD-IoU), into the RT-DETR model, enabling the identification of a wider range of flames and flame-like phenomena. Additionally, the use of Complete Local Binary Pattern (CLBP) technology for texture feature extraction, combined with VQGAN technology for accurate flame identification through sample reconstruction, underscores our innovative approach. The experimental results demonstrate the exceptional performance of the model, achieving precision, recall, F1 score, and accuracy rates of 98.23%, 96.33%, 97.33%, and 95.08%, respectively, on the Dataset for Fire and Smoke Detection (DFS), substantially surpassing existing methods. Our objective is to further develop FFDNet into a robust, efficient, and widely applicable tool for flame detection, thereby providing significant technical support for fire prevention and response initiatives.
火灾探测技术对于保障公共安全和减少财产损失至关重要。尽管传统方法和现代深度学习模型都取得了进展,但诸如精度欠佳和误报率高等挑战依然存在,尤其是在复杂的环境场景中。本文介绍了火灾聚焦检测网络(FFDNet),这是一种将经典方法与深度学习策略无缝集成的先进火焰检测框架。通过将增强的实时检测变压器(RT-DETR)模型与矢量量化生成对抗网络(VQGAN)相结合,我们的方法不仅提高了火焰检测的灵敏度和精度,还显著降低了误报频率。具体而言,我们在RT-DETR模型中集成了一种新颖的损失函数,即创新最小周长距离交并比(InnMPD-IoU),从而能够识别更广泛的火焰和类似火焰的现象。此外,使用完整局部二值模式(CLBP)技术进行纹理特征提取,并结合VQGAN技术通过样本重建进行准确的火焰识别,突出了我们的创新方法。实验结果表明该模型具有卓越的性能,在火灾和烟雾检测数据集(DFS)上分别实现了98.23%、96.33%、97.33%和95.08%的精度、召回率、F1分数和准确率,大大超过了现有方法。我们的目标是将FFDNet进一步发展成为一种强大、高效且广泛适用的火焰检测工具,从而为火灾预防和应对举措提供重要的技术支持。