| 杨明浩,依铭,何芳州.基于随机聚合深度激活图的火灾图像纹理检测[J].火灾科学,2026,35(1):87-94. |
| 基于随机聚合深度激活图的火灾图像纹理检测 |
| Fire image texture detection based on randomly aggregated deep activation maps |
| |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI：10.3969/j.issn.1004-5309.2026.01.10 |
| 基金项目:辽宁省社会科学规划基金项目(L21ASH004);辽宁省高校智库联盟专项课题项目(LJKZK-Y202301) |
|
| 中文关键词: 火灾监测 纹理识别 深度激活图 随机自编码器 |
| 英文关键词:Fire monitoring Texture recognition Deep activation maps Randomized autoencoder |
| 摘要点击次数: 27 |
| 全文下载次数: 0 |
| 中文摘要: |
| 火灾检测的准确性和实时性对社会安全至关重要,但现有方法在复杂环境中计算成本高、鲁棒性不足且需频繁微调。提出了一种轻量化、高精度的火灾纹理检测方法,将预训练卷积神经网络的深度激活图与随机自编码器(RAE)相结合,高效提取火灾图像中的纹理特征。实验结果表明,该方法在分类精度上显著优于传统的全局平均池化(GAP)方法。在ResNet18和ResNet50网络上,其准确率相比GAP分别提高了17.91%和9.73%;在ConvNeXt-T上,比GAP聚合版本(GAP agg)的准确率高出11.29%。同时,该方法可以显著降低计算成本,且无需模型微调,在复杂环境中具备强鲁棒性和广泛适用性,展示出在实时灾害监测领域的潜力。 |
| 英文摘要: |
| The accuracy and real-time performance of fire detection are crucial for social safety. However, existing methods are computationally expensive, lack robustness, and require frequent fine-tuning in complex environments. This study proposes a lightweight, high-precision fire texture detection method that combines the deep activation maps of pre-trained convolutional neural networks with a random autoencoder (RAE) to efficiently extract texture features from fire images. Experimental results show that this method significantly outperforms the traditional global average pooling (GAP) method in classification accuracy. On ResNet18 and ResNet50 networks, its accuracy is increased by 17.91% compared to GAP. On ConvNeXt-T, its accuracy is 11.29% higher than that of the GAP aggregation version (GAP agg). Moreover, this method significantly reduces computational costs, requires no model fine-tuning, demonstrates strong robustness and wide applicability in complex environments, and shows potential for real-time disaster monitoring. |
| 关闭 |
|
|
|
