Abstract:Nowadays, automatic fire alarm technology has been gradually advancing towards intelligence, networking, and automation. However, the current real-time flame detection techniques suffer from low detection accuracy and high computational parameters. To address these issues, this study focuses on the YOLOX-m object detection model and proposes an improved lightweight fire detection model. The proposed model improves the YOLOX-m model by replacing the CSPDarknet-53 backbone network with ShuffleNetV2, which reduces computational load while improving network accuracy. Additionally, the RFB module is inserted into the ShuffleNetV2 structure to increase the receptive field, enhancing the detection capability for large objects while maintaining resolution and precise localization. The upsampling in the Neck part is replaced with Pixel Shuffle to minimize feature loss. Furthermore, the CBAM attention mechanism is incorporated to enable the network to focus on crucial information, thus improving overall model performance. Through algorithm optimization and experimental testing, the improved model achieves a 2.87% increase in accuracy compared to the YOLOX-m model, with a reduction of 37.9% in parameters and a 30.7% decrease in computational load. The improved model has been successfully applied to real-world scenarios such as forest fires and urban fires. By comparison, flames can be detected relatively more accurately.