基于改进YOLOv5s的焦炉烟火识别算法

针对炼焦厂烟火排放全天候环保监测的要求,提出了基于改进YOLOv5s的焦炉烟火识别算法;该算法以YOLOv5s为基础网络,在主干网络Backbone中添加CBAM注意力机制模块,使网络更加关注重要的特征,提升目标检测的准确率;新增FReLU激活函数代替SiLU激活函数,提高激活空间的灵敏度,改善烟火图像视觉任务;在自建数据集中烟、火样本标签基础上,增加灯光标签来解决强灯光对火焰识别的干扰,并通过分流训练、检测的方式来解决昼夜场景的烟火检测问题;在自建数据集上做对比实验,更换激活函数后,联合CBAM模块的YOLOv5s模型效果最佳;实验结果显示,与原始YOLOv5s模型相比,在白天场景下的烟火识别mAP值提升了6.7%,在夜间场景下的烟火识别mAP值高达97.4%。

Coke Oven Smoke and Fire Recognition Algorithm Based on Improved YOLOv5s

For the requirements of all-weather environmental monitoring of smoke and fire emissions from coke plants, a coke oven smoke and fire recognition algorithm based on improved YOLOv5s is proposed; the algorithm uses YOLOv5s as the base network and adds CBAM"s attention mechanism module to the reference network, so that the network focuses more on relevant features and improves target detection accuracy; a new FReLU activation function replaces the SiLU activation function to improve the sensitivity of the activation space and improve the smoke and fire image vision task; on the basis of smoke and fire sample labels in the self-built dataset, add light labels to solve the interference of strong lights on flame recognition, and solve the smoke and fire detection problem of day and night scenes by shunting training and detection; do comparison experiments on the self-built dataset, after replacing the activation function, the joint CBAM module of the experimental results show, that the mAP value of smoke and fire detection in the day scene is improved by 6.7% compared with the original YOLOv5s model, and the mAP value of smoke and fire recognition in nighttime scenes is as high as 97.4%.