一种基于改进 TransUnet 的火焰分割模型

A Flame Segmentation Model Based on Improved TransUnet

摘要

近年来火灾事故频发，严重威胁到全球生态、经济及人类生命安全。在火灾检测方面，基于传感器的传统火灾检测方法在覆盖范围和误报率方面存在局限，难以应对复杂多变的火灾环境。为此，本文提出了一种基于 TransUnet 的改进火焰分割模型，该模型融合了多尺度边界注意力网络与增强双重注意力模块，多尺度边界注意力网络利用 Sobel 算子和多尺度特征融合，提升了模型对火焰边界的分割准确率；增强双重注意力模块通过 Sim AM、位置注意力和通道注意力，增强了局部细节的提取能力，优化模型对火焰细节特征的提取。在 Flame 和 BowFire 数据集上验证，实验证明该模型提升了火焰分割的精度，在复杂情况下有着较好的鲁棒性，为火灾检测提供了更加精准的解决方案。

Abstract

In recent years，frequent fire accidents have seriously threatened the global ecology，economy and hu-man life safety.In terms of fire detection，traditional sensor-based fire detection methods have limitations in cover-age and false alarm rate，and are difficult to cope with complex and changeable fire environments.Therefore，an improved flame segmentation model based on TransUnet is proposed in this paper.The model integrates the multi-scale boundary attention network and the enhanced dual attention module.The multi-scale boundary attention net-work uses Sobel operator and multi-scale feature fusion to improve the segmentation accuracy of the flame bound-ary.The dual attention module enhances the ability to extract local details through Sim AM，position attention and channel attention，and optimizes the model＇s extraction of flame detail features.Experiments on Flame and BowFire datasets show that the model improves the accuracy of flame segmentation and has good robustness in complex situations，providing a more accurate solution for fire detection.

关键词

深度学习 / 图像分割 / 火灾检测 / 注意力机制 / TransUnet

Key words

deep learning / image segmentation / fire detection / attention mechanism / TransUnet

参考文献

[1] Marjani M, Ahmadi S A, Mahdianpari M. FirePred:A hybrid multi-temporal convolutional neural network model for wildfire spread prediction[J]. Ecological In-formatics, 2023, 78: 102282.

[2] 李倬毅, 孟骏, 杨晓冬, 等. 基于 CNN-LSTM 模型的燃烧烟气 CO₂ 浓度预测研究[J]. 燃烧科学与技术, 2025, 31 (4):406-414.

[3] Li Zhuoyi, Meng Jun, Yang Xiaodong, et al. Predic-tion of flue gas CO₂ concentrations based on the CNN-LSTM mode[J]. Journal of Combustion Science and Technology, 2025, 31 (4):406-414 (in Chinese).

[4] 顾成杰, 高紫莲, 朱东郡, 等. 基于跨尺度特征增强与多层注意力机制的火灾检测方法[J]. 燃烧科学与技术, 2026, 32 (1):95-108.

[5] Gu Chengjie, Gao Zilian, Zhu Dongjun, et al. Fire de-tection method based on cross-scale feature enhancement and multi-layer attention mechanism[J]. Journal of Combustion Science and Technology, 2026, 32(1): 95-108 (in Chinese).

[6] 陈跨越, 王保云. 基于改进 Resnet18 网络的火灾图像识别[J]. 河南师范大学学报(自然科学版), 2024, 52 (4):101-112.

[7] Chen Kuayue, Wang Baoyun. Fire image recognition based on improved Resnet18 network[J]. Journal of He-nan Normal University(Natural Science Edition), 2024, 52 (4):101-112 (in Chinese).

[8] Cho Ah Young, Si-eun Park, Duk-jin Kim, et al. Burned area mapping using Unitemporal Planet Scope imagery with a deep learning based approach[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2022, 99:242-253.

[9] Roy David P, Haiyan Huang, Rasmus Houborg, et al. A global analysis of the temporal availability of Planet-Scope high spatial-resolution multi-spectral imagery[J]. Remote Sensing of Environment, 2021, 264: 112586.

[10] Randerson J T, van der Werf G R, Giglio L, et al.Global Fire Emissions Database,Version 4.1 (GFEDv4) [EB/OL].https://doi.org/10.3334/ORNLDAAC/1293,2025-04-24.

[11] 朱莉, 赵俊, 傅应锴, 等. 一种红外热图像目标区域分割的深度学习算法[J]. 西安电子科技大学学报, 2019, 46(4):107-114, 121.

[12] Zhu Li, Zhao Jun, Fu Yingkai, et al. Deep learning algorithm for the segmentation of the interested region of an infrared thermal image[J]. Journal of Xidian Univer-sity, 2019, 46 (4):107-114, 121 (in Chinese).

[13] Seydi S T, Hasanlou M, Chanussot J. Burnt-Net: Wildfire burned area mapping with single post-fire Senti-nel-2 data and deep learning morphological neural net-work[J]. Ecological Indicators, 2022, 140: 108999.

[14] 戴洋毅, 何康, 瑚琦, 等. CNN-Transformer 混合模型在计算机视觉领域的研究综述[J]. 建模与仿真, 2023, 12(4):3657-3672.

[15] Dai Yangyi, He Kang, Hu Qi, et al. Review of CNN-Transformer hybrid model in computer vision[J]. Model-ing and Simulation, 2023, 12(4):3657-3672( in Chi-nese).

[16] Ghali R. Wildfire segmentation using deep vision trans-formers[J]. Remote Sensing, 2021, 13 (17): 3527.

[17] Dosovitskiy Alexey, Beyer Lucas, et al. An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale[EB/OL]. https://arxiv.org/abs/2010.11929,2020-10-22.

[18] Wang G, Wang F, Zhou H, et al. Fire in focus: Advancing wildfire image segmentation by focusing on fire edges[J]. Forests, 2024, 15(1): 217.

[19] Jing Tao, Meng Qinghao, Hou Hongru. SmokeSe-ger:A transformer-CNN coupled model for urban scene smoke segmentation[J]. IEEE Transactions on Industrial Informatics, 2023, 20 (2):1385-1396.

[20] Chen X, Qi D, Shen J. Boundary-Aware Network for Fast and High-Accuracy Portrait Segmentation[EB/OL].

[21] Shamsoshoara A, Afghah F, Razi A, et al. Aerial im-agery pile burn detection using deep learning:The FLAME dataset[J]. Computer Networks, 2021, 193: 108001.

[22] Chino Daniel Y T, Letricia P S Avalhais, Jose F Rodri-gues, et al. Bowfire:Detection of fire in still images by integrating pixel color an-d texture analysis[C]// 2015 28th SIBGRAPI Conference on Graphics,Patterns and Images.Salvador,Brazil:IEEE, 2015:95-102.

基金资助

国家自然科学基金资助项目(31771224)