Abstract
This paper proposes modifications to the YOLOv8 architecture for visual smoke detection in early fire detection systems. To reduce the complexity of the architecture, it is proposed to use the VoVGSCSP block, to improve the detection quality – the efficient multi-scale attention (EMA) block. The proposed changes allowed to reduce the number of neural network parameters by 27% and the computational complexity (GFLOPS) by 12% compared to the YOLOv8n model. The proposed model retained high detection accuracy, compared to the base model, the decrease in detection accuracy by the quality assessment metrics mAP@0.5 and mAP@0.5:0.95 was about 1%. To train and evaluate the model, an own dataset of more than 5000 images was created based on the open datasets D-Fire and WSDY. The obtained results demonstrate the suitability of the model for use on edge devices as part of video surveillance systems for early fire detection.