LWSCA-YOLOV5: An Improved Lightweight Fire Detection Algorithm Based on YOLOV5

To address the challenges posed by parametric and computationally intensive deep learning-based fire detection methods, hindering real-time detection solutions, we propose a lightweight fire detection algorithm named LWSCA-YOLOv5. Our method achieves strong detection performance with a significant reduction in FLOPs and Params. Firstly, we introduce a novel lightweight convolution module called SCAConvolution. This operation enables the embedding of object position information into the channel dimension during the feature fusion process. In comparison to standard convolution, SCAConvolution substantially reduces computational consumption while extracting richer features. Subsequently, we leverage this convolution unit to redesign the C3 module in YOLOv5, resulting in the proposed LWC3. To fully harness the potential of our method, we introduce the SPPAM module. Finally, we assemble the lightweight fire detection network using the aforementioned components, naming it LWSCA-YOLOv5. Experimental results demonstrate a 71% reduction in Params and a 63% reduction in FLOPs compared to the baseline while maintaining the same level of accuracy, validating the effectiveness of our method.