基于 YOLOv5的多传感器可迁移林火检测模型

A Multi-sensor Forest Fire Detection Model Based on YOLOv5

摘要:

快速准确地检测森林火灾中的火点对于减轻火灾损失和开展有效的灭火救援工作至关重要 。深度学习技术可 以 自动学习并提取不同传感器获取的林火特征 。无人机搭载不同类型的传感器可以快速获取森林实时图像,但不同传感 器获取的图像特征不同,导致深度学习模型在不同图像中的迁移性较差 。本研究基于 YOLOv5预训练模型,使用 286张可 见光和红外图像进行迁移学习,以增强模型对不同图像类型的适用性 。迁移学习后,模型对可见光图像测试集的检测精度 提升了 6%,对红外图像测试集的检测精度达到 0. 952,证明模型对两种传感器图像数据均具有较强的检测能力 。相比可 见光模型,多模态模型在包含可见光和红外图像的测试集上的mAP50达到 0. 914,表明模型成功地提高了对红外图像的检 测能力,并保留了对可见光图像的良好检测性能 。综上,本研究采用少量图像进行迁移学习 ,取得了较好结果 。迁移学习 可以使 YOLOv5火灾检测模型适配红外图像,在保留可见光图像检测能力的同时,增强其对火情的检测准确性与环境适应 性 。这为实现基于多源数据的森林火灾检测奠定了基础,为提高森林火灾检测系统的智能化与 自适应提供了思路。

Abstract:

Rapid and accurate detection of fire spots in forest fires is crucial for reducing fire losses and conducting effec- tive firefighting and rescue operations . Deep learning techniques can automatically learn and extract forest fire features obtained from different sensors . unmanned aerial vehicles equipped with different types of sensors can quickly obtain real —time images of forests, but the different image features acquired by different sensors lead to poor transferability of deep learning models in different images . In this study, based on the yOLOv5pre—trained model for fire detection, we used 286 visible light and infrared images for transfer learning to enhance the model》s applicability to different types of images . Af- ter transfer learning, the model achieved a detection accuracy of 0. 761for the visible light image test set and 0. 952for the infrared image test set. This demonstrates that the model has strong detection capabilities for both types of sensor image data . compared with the original model, the mAP50of the model after transfer learning on the test set containing visible light and infrared images reached 0. 914. This indicates that the model successfully improved its detection ability for infra- red images while preserving good detection performance for visible light images . In summary, this study used a small num- ber of images for transfer learning and achieved good results . Transfer learning can adapt the yOLOv5fire detection model to infrared images,enhancing its detection accuracy and environmental adaptability for fire situations while retaining the detection capabilities for visible light images . This lays the foundation for realizing multi—source data—based forest fire detection and pro- vides ideas for improving the intelligence and adaptability of forest fire detection systems .