基于YOLOv5的改进火灾检测算法

Fire Detection Algorithm Based On Improved YOLOv5

针对火灾场景下现有目标检测算法对起火目标的漏检误检问题,提出一种基于改进YOLOv5的火灾检测算法。该算法引入Ghost-Atrous模块,能够在不增加卷积核大小和数量的情况下扩大感受野,从而替代普通“卷积+池化层”的组合,以避免特征信息丢失,同时减少参数量和运算量;使用CBAM-Atrous注意力模块强化对重要特征的提取;采用EIOU-NMS进行非极大值抑制,更好地解决误检漏检问题。实验结果表明,改进后的算法在火灾数据集上相比于原始YOLOv5算法mAP@0.5提高了4.3个百分点;相比于其他主流的目标检测算法,mAP@0.5提高了0.3~8.0个百分点,同样具有一定的优越性。

Aiming at the problem of missing detection and misdetection of fire targets by existing target detection algorithms in fire scenes,a fire detection algorithm based on improved YOLOv5 is proposed.The algorithm introduces the Ghost-Atrous module,which can expand the receptive field without increasing the size and number of convolution nuclei,so as to replace the combination of common convolution+pooling layers to reduce the loss of feature information,and reduce the number of parameters and calculation.The CBAM-Atrous attention module is used to enhance the extraction of important features.EIOU-NMS is used for non-maximum suppression to better solve the problem of false detection and leakage detection.The experimental results show that the improved algorithm outperforms the original YOLOv5 algorithm on fire datasets mAP@0.5 Increased by 4.3 percentage points;Compared to other mainstream object detection algorithms,mAP@0.5 Improved by 0.3-8.0 percentage points,also possessing certain advantages.