基于改进YOLOv5算法的火灾图像检测研究

Research on Fire Image Detection Based on Improved YOLOv5 Algorithm

针对传统的火灾检测系统检测时间长、容易误报等问题,提出了一种基于改进YOLOv5的火灾检测算法.首先,针对卷积操作受到感受野的限制,增加Involution算子来扩大感受野;其次,在模型中引入CA(coordinate attention)注意力机制,重新分配特征图高度和宽度上的权重;最后,使用定位损失函数α-CIoU替换CIoU,以提高模型回归精度.改进的YOLOv5算法的平均精度达到了68.4%,相较于标准算法的平均精度提高了3.9%,定位框更加准确,且检测速度提高了14%.实验结果表明改进后的算法模型明显提高了火灾检测的准确性和实时性.

Aiming at the problems of traditional fire detection system,such as long detection time and easy false alarm,this paper proposes a fire detection algorithm based on improved YOLOv5.Firstly,because convolution operation is limited by receptive field,the convolution operator is added to expand receptive field.Secondly,CA(coordinate attention)mechanism is introduced into the model to redistribute the weight of the height and width of the feature map.Finally,the localization loss functionα-CIoU is used to replace CIoU to improve the regression accuracy of the model.The average accuracy of the improved YOLOv5 algorithm is increased by 68.4%.Compared with the standard algorithm,the average accuracy is increased by 3.9%,the positioning frame is more accurate and the detection speed is increased by 14%.The experimental results show that the improved algorithm model obviously improves the accuracy and realtime performance of fire detection.