RCM: A Novel Fire Detection Technique That Effectively Resists Interference in Complex Scenarios

Fire detection is critical to life. However, the reality of the scene is very complex. Interference from various scenarios makes it difficult to detect fires in a timely manner, e.g., there may be occluded fires in complex scenarios, which results in smaller flame targets that are difficult to detect and higher leakage and false detection rates. To address the above limitations, we develop a novel deep learning method, i.e., RCM, for complex scenarios. Specifically, it expands the model’s sensory field to capture the fire background information more extensively by designing a novel residual feature extraction network, i.e., RDNet, and combines it with the convolutional block attention module (CBAM) so that the model can pay more attention to the fire region, thus effectively reducing the interference of a complex background. Then, the multiscale feature fusion (MFF) module provides the model with a multilevel feature representation, which enhances the model’s ability to capture detailed features of smaller fires. Moreover, we introduce a new loss function, MPDIoU, to improve the convergence speed and detection accuracy of the model. Experiments on real fire scene datasets show that the precision and accuracy of the RCM model reach 94.77% and 97.99%, respectively, while the FPR and FNR are only 1.19% and 5.57%. Meanwhile, the detection performance of the method is optimized compared to more advanced algorithms in the field. The method demonstrates the effectiveness of fire detection in complex scenes and provides a novel and effective technique for fire detection.