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.
deep learningYOLOv5 algorithmfire detectionGhost-Atrousatrous convolution
万方数据
维普
