Wildfire smoke detection based on enhanced YOLOv7 with video images

Abstract

An improved YOLOv7-I-CBAMI model (enhanced You Only Look Once version7 integrated with the Convolutional Block Attention Module) combined with ridge edge detection is proposed to detect wildfire smoke in mountainous areas. YOLO (You Only Look Once) reformulates object detection as a regression task, utilizing the entire image as input and generating bounding box coordinates and class labels through a single neural network, offering high detection accuracy and speed. However, limitations exist in detecting closely spaced and small objects, and distinguishing between clouds and wildfire smoke remains an unresolved issue. To address these challenges, a bidirectional feature pyramid network is introduced to improve detection accuracy, and an enhanced CBAM (Convolutional Block Attention Module) attention mechanism is incorporated to overcome YOLOv7′s limitations in detecting small targets and faint wildfire smoke features. Furthermore, ridge edge detection is integrated for secondary optimization, reducing the confusion between wildfire smoke and natural clouds. Experimental results on a wildfire-prone transmission corridor video dataset around Kunming, provided by Yunnan Power Grid, indicate that the YOLOv7-I-CBAMI network achieves superior performance in Precision, Recall, and F1-Score. By integrating ridge edge detection with wildfire smoke detection, Precision is improved and overall performance metrics are also enhanced, achieving final values of 0.83 for Precision, 0.82 for Recall, and 0.82 for F1-Score, with a detection speed of 21.20 FPS (Frames Per Second). These results validate the effectiveness of the proposed YOLOv7-I-CBAMI model with ridge detection for rapid and accurate detection of wildfire smoke in transmission corridors.