A lightweight forest smoke detection algorithm

Forest fires have always been a global issue, and the prevention and control of forest fires have always been a worldwide challenge. Traditional forest smoke detection methods have problems such as small detection range, high false positive rate, and low detection accuracy. In recent years, computer vision technology has been applied to forest fire smoke detection. However, existing models still face challenges in complex environments. This paper proposes a lightweight early forest fire smoke detection model based on partial convolution. The backbone network mainly consists of the DIRPC module responsible for downsampling and the IRPC module responsible for feature extraction. Both modules adopt the inverted residual structure and partial convolution operation, which effectively deepen the network layers while reducing the number of parameters and GFLOPs. A lightweight feature fusion network of SlimNeck is adopted, with YOLOv8 Head as the detection head, constructing a one-stage early forest fire smoke detection model. Experiments on datasets of various forest scenes show that our network has better detection accuracy and fewer parameters and GFLOPs compared to other mainstream object detection networks.