Research on the Detection Method for Early Fire Based on YOLOv5m Optimization Model

Due to the extensive application of vision in fir e monitoring, the optimization method of the YOLOV5m mo del is researched for detecting the early fire accidents accor ding to the flame and smoke of early fire are taken as the detection targets. Sa attention mechanism is introduced into t he YOLOV5m model backbone network to improve the lear ning ability of the model to the characteristics of flame and smoke, and then suppress the background interference infor mation meanwhile. Furthermore, the standard convolution of the original model is replaced by the deep separable convol ution, and the model is lightweighted while maintaining the model detection performance. By replacing the CIOU loss fu nction of Yolov5m with the alpha-EIOU loss function, the q uality and regression precision of the final prediction frame work for flame and smoke targets are improved. The superi ority of the optimization method is verified by the comparis on experiment of YOLOV5m model before and after optimiz ation, and the comparison experiment of the optimized mode 1 with other typical models.