Abstract:Aiming at the trouble of low detection accuracy and the problem of large model size, this paper proposes a lightweight flame-and-smoke detection model depending on global awareness of images. The proposed method replaces the Conv+BatchNorm+SiLU (CBS) module of original you only look once version 5 (YOLOv5) in the backbone with DSConv+BatchNorm+SiLU (DBS), and the C3 module with GC3, and thus constructs a lightweight backbone network. Besides, involution (InvC3) module is proposed to enhance the global modeling ability and compress the model size, and a module using adaptive receptive fields, named FConv, is proposed to enhance the model’s perception capacity for foreground complex flame-and-smoke information in feature maps. Experimental results show that the proposed model increases the mean average precision of all categories at 0.5 IOU (mAP@0.5) to 70.8%, the mAP@0.5:0.95 to 39.7%, reduces the number of parameters to 3.57M and the amount of calculation to 7.4 giga floating-point operations per second (GFLOPs) under the premise of ensuring the detection speed. It has been verified that the model can achieve high-precision real-time detection of flame and smoke.