基于YOLOv3的自注意力烟火检测算法

Self-attention Smoke and Fire Detection Based on YOLOv3

烟火图像检测在火灾防控中具有重要的意义,但由于烟雾和火焰成像具有多变性和无规则性,大多数烟雾检测算法在现实场景下往往表现欠佳,漏检情况相对严重。针对于此,通过结合基于深度学习的一阶段回归目标检测模型(YOLOv3)与自注意力机制,提出了一种基于YOLOv3自注意力烟雾和火焰图像检测算法,在优化原有YOLOv3网络模型的基础上,通过结合多尺度的自注意力网络,融合模型上下文信息,引导模型自适应学习提取关键的特征信息,增强了模型的特征表达,可有效提高烟雾和火焰检测的准确率。在测试集上结果表明,相比原有的基准检测算法,该方法烟火检测算法的准确率为92.1%,提高了6.5%;通过主干网络替换对比实验进一步验证了该结构设计的有效性,该算法具备较好的实用性。

Pyrotechnic image detection has great significance in fire prevention and control.As the variability and irregularity of smoke and fire imaging,most of the detection algorithms perform poorly in real scenes.By combining the one-stage regression object detection based on deep learning(YOLOv3)and the self-attention mechanism,a self-attention smoke and fire detection was proposed based on YOLOv3.With the multi-scale self-attention network,which fusing context guides adaptive learning of the modeland extracts key information from features,model can effectively enhance the feature expressionand improve the accuracy of smoke and flame detection.Results show the accuracy of smoke and fire detection has significantly improved.The accuracy of the fireworks detection method in this paper is 92.1%,which is an increase of 6.5%.The comparison experiment of backbone network replacement further verifies the effectiveness of the structure design,and the algorithm has good practicability.