A Smoke Detection Model Based on Improved YOLOv5

Fast and accurate smoke detection is very important for reducing fire damage. Due to the complexity changeable nature of scenes, existing technology has problems a low rate high false negative rate, robustness generalization ability algorithms are not high. Therefore, this paper proposes model based on improved YOLOv5. First, large number real synthetic images were collected form dataset. Different loss functions (GIoU, DIoU, CIoU) used three different models YOLOv5 (YOLOv5s, YOLOv5m, YOLOv5l), YOLOv5m was as baseline model. Then, because problem small numbers training samples, mosaic enhancement method randomly crop, scale arrange nine new images. To solve inaccurate anchor box prior information in YOLOv5, dynamic mechanism proposed. An generated dataset through k-means++ clustering algorithm. The module added model, size position dynamically updated network process. Aiming at unbalanced feature maps scales an attention proposed improve performance by adding channel spatial original structure. Compared with traditional deep learning algorithm, algorithm 4.4% higher than mAP speed reached 85 FPS, which obviously better can meet engineering application requirements.