一种改进YOLOX_S的火焰烟雾检测算法

Flame Smoke Detection Algorithm Based on Improved YOLOX_S

针对目前在火灾预警方面还存在火焰烟雾检测效果差、误报率高等问题,在YOLOX框架下提出改进YOLOX_S目标检测算法。首先在数据集建立方面,采用的数据集包括Bilkent University公开的数据集和部分自建数据集,共计9621张图片。并且通过对数据集采用Mosaic数据增强的方式,增加数据的多样性。其次对backbone部分采用swin-T骨干网络来代替原来的CSPDarkNet骨干网络,能够更好的捕捉不同尺度下的特征,有效地提升了目标检测的精度。然后对网络模型引入加权双向特征金字塔网络(bidirectional feature pyramid network,BiFPN)特征融合网络,提高检测的效率和网络模型的适应性,在复杂背景下同样可以保持较高的检测精度。最后引入CA注意力机制来加强此算法的特征提取能力。经过对比实验表明,改进后的YOLOX_S的火焰烟雾检测算法具有较高准确性,其mAP@0.5(预测框与真实框重合程度的阈值为0.5时的平均检测精度)达到81.5%,相比原网络提高了5.3%。改进后的YOLOX_S网络模型在火焰烟雾检测方面具有更高准确性和更低的误报率。

In response to the current issues of poor flame and smoke detection performance and high false alarm rates in fire warning,an improved fire and smoke detection algorithm was proposed based on the YOLOX framework.In the data collection phase,the used dataset consists of 9621 images,including data from the publicly available Bilkent University dataset and a portion of self-built data.The data diversity was enhanced through the adoption of Mosaic data augmentation.Subsequently,the original CSPDarkNet backbone was replaced with the swin-T backbone network,which better captured features at different scales,effectively improving the accuracy of object detection.Additionally,the BiFPN(bidirectional feature pyramid network)feature fusion network was introduced to the network model,enhancing detection efficiency and adaptability,thus maintaining high detection accuracy even in complex backgrounds.Finally,the CA attention mechanism was incorporated to strengthen the feature extraction capabilities of this algorithm.Comparative experiments show that the improved YOLOX_S fire and smoke detection algorithm achieves high accuracy,with mAP@0.5(the average detection accuracy when the threshold for the degree of overlap between the predicted box and the true box is 0.5)reaching 81.5%,representing a 5.3%improvement compared to the original network.The improved YOLOX_S network model has higher accuracy and lower false alarm rate for flame smoke detection.