Research on YOLOv5 forest fire recognition algorithm utilizing attention mechanism

In recent years, various fire incidents have become increasingly prevalent, posing a severe threat to both life and property safety. The timely detection, early warning, and prevention of fire smoke are crucial for minimizing losses and ensuring sustainable societal development. Traditional forest fire recognition algorithms are significantly impacted by environmental conditions, making them unsuitable for expansive regions. The majority of image-based algorithms rely on manual feature extraction, contingent on the expertise of researchers. Deep learning, leveraging abundant input samples, automatically extracts features through multiple layers of neural nodes, resulting in a model with superior generalization capabilities. This paper introduces a forest fire detection algorithm based on convolutional neural networks by incorporating the non-parametric attention mechanism SimAM into the core network of YOLOv5. The attention mechanism assigns three-dimensional attention weights to wildfire feature maps, thereby augmenting the model's feature capture proficiency. Experimental results show that, without introducing additional parameters, compared to the original YOLOv5 algorithm, the accuracy, recall, and F1-score increased by 9.6%, 8.2%, and 6.2%, respectively, achieving precise identification of forest fires.