ECS-STPM: An Efficient Model for Tunnel Fire Anomaly Detection

The fire spreads rapidly in the tunnel due to the narrow space and high sealing, which makes rescue hard and threatens the citizen’s lives. However, the lack of public fire datasets makes it challenging for networks to learn targeted representations of fire features, resulting in low detection accuracy. To tackle this problem, we construct a Tunnel Fire Anomaly Detection (TF-AD) dataset based on unsupervised training. This dataset contains 5200 high-resolution color images, including non-fire images for training and fire images with annotations for testing. Based on the TF-AD dataset, we propose an efficient tunnel fire anomaly detection model named ECS-STPM. ECS-STPM consists of a teacher and student network with identical EfficientNet-B1 structures. Additionally, considering the efficiency of adaptively assigning channel weights, we combine the convolutional kernel with channels to propose a novel attention mechanism, Efficient Kernel and Channel Attention (EKCA). EKCA replaces the Squeeze-and-Excitation (SE) networks in the MBConv module to prevent the loss of crucial information. Furthermore, we introduce the SPD-Conv module instead of the strided convolution layer to increase the detection accuracy in smaller fire areas. The experimental results on TF-AD dataset show that the pixel-level AUC-ROC and image-level AUC-ROC are up to 0.931 and 0.835, which verifies the effectiveness of our model.