Recognition of Forest Fire Smoke Based on Improved YOLOv8n Model

Abstract

To address the challenges of early forest fire smoke image recognition, including false alarms and missed reports caused by interference in complex environments, an enhanced model, named MB-YOLO, is proposed based on the YOLOv8 Nano (YOLOv8n) architecture for efficiently recognizing forest fire smoke. Firstly, to overcome detection failures of low-concentration smoke in complex backgrounds, the original Path Aggregation Network (PAN) is replaced with a bi-directional feature pyramid network (BiFPN). This substitution not only enhances multi-scale feature extraction but also simplifies the network structure, reducing the number of parameters. Secondly, to address false detections caused by cloud and mist interference, the C2f_MLCA module is developed. This module integrates a lightweight Mixed Local Attention mechanism (MLCA) into the bottleneck of the gradient flow module C2f, thereby enhancing smoke feature extraction. Lastly, to reduce sensitivity to positional offsets of small smoke targets, the Complete Intersection over Union (CIoU) loss is replaced with Inner-DIoU loss. This new loss function computes loss with auxiliary bounding boxes, accelerating convergence speed and enhancing accuracy for small smoke targets. The effectiveness of the algorithm is validated with a curated dataset containing small smoke targets, unclear backlighting, and cloud and mist interference. Experimental results demonstrate that our model achieves a mean Average Precision (mAP) of 80.1%, a frame rate of 60.6 Frames Per Second (FPS), with a total of 1.09 million parameters and 7.1 billion floating-point operations per second (FLOPs). This model offers high detection accuracy, fewer parameters, and lower GFLOPs, facilitating accurate real-time monitoring of forest fires in complex environments and all-weather conditions.