| 基于改进YOLOv8 的火灾图像分类方法 |
| 点此下载全文 |
| 引用本文：朱定赟,陈翼遥.基于改进YOLOv8 的火灾图像分类方法[J].上海第二工业大学(中文版),2025,42(1):59-65 |
| 摘要点击次数: 252 |
| 全文下载次数: 40 |
|
|
| 中文摘要:鉴于传统消防报警系统在火灾环境下难以有效区分烟雾与火焰, 导致响应延迟或误报的问题, 提出了一种火灾图像识别方法。该方法基于深度迁移学习技术并对You Only Look Once v8 (YOLOv8) 网络模型进行了优化。具体而言, 在模型中引入了瓶颈注意力模块(bottleneck attention module, BAM) 以增强特征提取网络的性能, 显著提升模型对火灾特征(如烟雾与火焰) 的敏感度和区分能力。同时为了克服内部协变量偏移的问题, 应用了批量通道归一化(batch channel normalization, BCN) 技术于每一层数据输出阶段, 有效降低了因线性变换及激活函数作用导致数据范围扩大的影响。在广泛使用的D-Fire 火灾数据集上进行的一系列严格实验表明, 优化后的模型展现出卓越的性能。相较于原始YOLOv8 模型, 改进后的版本在精确性、全面捕获火灾事件的能力(召回率) 以及综合评估指标(F1 得分) 上均实现了显著提升。 |
| 中文关键词:图像分类 You Only Look Once v8 (YOLOv8) 火焰检测 深度学习 |
| |
| Fire Image Classification Method Based on Improved YOLOv8 |
|
|
| Abstract:In view of the fact that the traditional fire alarm system is difficult to effectively distinguish between smoke and flame in a fire environment, which leads to the problem of delayed response or false alarm, an innovative fire image recognition method is proposed, which is based on deep transfer learning technology and optimized by You Only Look Once v8 (YOLOv8) network model. Specifically,the bottleneck attention module (BAM) is introduced to enhance the performance of the feature extraction network, which significantly improves the sensitivity and discrimination ability of the model for fire features (such as smoke and flame). At the same time, in order to overcome the problem of internal covariate shift, batch channel normalization (BCN) is used to the data output stage of each layer, which effectively reduces the impact of data range expansion caused by linear transformation and activation function. A series of rigorous experiments on the widely used D-Fire data sets show that the optimized model shows excellent performance. Compared with the original YOLOv8 model, the improved version achieves a significant improvement in accuracy, the ability to comprehensively capture fire events (recall rate), and the comprehensive evaluation index (F1 scores). |
| keywords:image classification You Only Look Once v8 (YOLOv8) flame detection deep learning |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|