| 胡久松,刘张驰,余谦,谷志茹,钟皓.融入GhostNet和CBAM的YOLOv8烟雾识别算法[J].电子测量与仪器学报,2024,38(8):201-207 |
| 融入GhostNet和CBAM的YOLOv8烟雾识别算法 |
| YOLOv8 smoke detection algorithm integrated with GhostNet and CBAM |
| |
| DOI： |
| 中文关键词: 早期火灾探测 烟雾识别 YOLOv8 GhostNet CBAM |
| 英文关键词:early fire detection smoke recognition YOLOv8 GhostNet CBAM |
| 基金项目:湖南省教育厅优秀青年项目（293832）、湖南省自然科学基金项目（2023JJ50198，2022JJ50005）资助 |
|
|
| Author | Institution |
| Hu Jiusong | College of Railway Transportation, Hunan University of Technology, Zhuzhou 412007, China |
| Liu Zhangchi | College of Railway Transportation, Hunan University of Technology, Zhuzhou 412007, China |
| Yu Qian | School of Marine engineering, Hunan Automotive Engineering Vocational University, Zhuzhou 412000, China |
| Gu Zhiru | College of Railway Transportation, Hunan University of Technology, Zhuzhou 412007, China |
| Zhong Hao | College of Railway Transportation, Hunan University of Technology, Zhuzhou 412007, China |
|
| 摘要点击次数: 33 |
| 全文下载次数: 92 |
| 中文摘要: |
| 火灾防控在公共安全中至关重要，传统探测手段在特定环境下面临局限。计算机视觉技术能实时监控大范围，精准识别烟雾等火灾前兆。但烟雾形状、纹理和颜色的复杂性给机器视觉的烟雾精准识别带来了极大挑战。针对这个问题，设计了一种融入轻量级网络和卷积块注意力机制的YOLOv8烟雾分类算法，旨在提升烟雾分类的精度与效率。首先，算法采用了GhostNet架构，通过替换传统的卷积层，保持高性能的同时，极大减轻了模型的负担。其次，算法嵌入了CBAM注意力机制，能够自动调整对不同区域的关注程度，确保关键烟雾特征被优先处理和精细分析，增强了模型的鲁棒性。采用公开烟雾数据集和加入挑战样本的自制数据集进行了大量实验。实验结果证明，算法烟雾识别准确率在公开数据集上达到了99.9%，在自制数据集达到了99.2%，优于同类方法。在实验电脑上，算法在GPU条件下帧率达到了833 fps，CPU条件下帧率达到了28 fps，可以用于快速准确地进行早期火灾探测。 |
| 英文摘要: |
| In the crucible of public safety, the imperative to guard against the scourge of fire is non-negotiable, yet conventional detection methodologies often falter when confronted with the complexities of specific environments. Herein lies the promise of computer vision technology, which offers the capability to monitor expansive territories in real-time and to identify the telltale signs of impending fires, most notably smoke. However, the intricate morphologies, textural variations, and chromatic subtleties of smoke present a significant challenge to the precision of its detection through machine vision.Addressing this exigency, we have conceived and developed an innovative smoke classification algorithm, seamlessly integrating a lightweight neural network and the convolutional block attention module (CBAM) within the YOLOv8 framework. This approach is designed to augment the accuracy and efficiency of smoke classification. Our algorithm leverages the GhostNet architecture, ingeniously replacing standard convolutional layers with a more efficient alternative, thereby maintaining high performance while drastically reducing the computational load on the model.Furthermore, the integration of CBAM imbues the algorithm with the ability to dynamically adjust its focus across different regions of the image, ensuring that salient smoke features are prioritized for detailed analysis. This feature enhances the model’s robustness and adaptability to diverse scenarios.To validate the efficacy of our algorithm, we conducted extensive experiments using both a publicly available smoke dataset and a custom dataset augmented with challenging samples. Empirical results have demonstrated that our algorithm achieves a smoke recognition accuracy of 99.9% on the public dataset and 99.2% on the custom dataset, outperforming existing methods. On our experimental machine, the algorithm exhibited a frame rate of 833 fps under GPU-accelerated conditions and 28 fps under CPU-only operation, affirming its potential for rapid and accurate early fire detection. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
