Hybrid CBAM-EfficientNetV2 Fire Image Recognition Method with Label Smoothing in Detecting Tiny Targets

Image fire recognition is of great significance in fire prevention and loss reduction through early fire detection and warning. Aiming at the problems of low accuracy of existing fire recognition and high error rate of tiny target detection, this study proposed a fire recognition model based on a channel space attention mechanism. First, the convolutional block attention module (CBAM) is introduced into the first and last convolutional layers EfficientNetV2, which shows strong feature extraction ability and high computational efficiency as the backbone network. In terms of channel and space aspects, the weights in the feature layer are increased, which enhances the semantic information of flame smoke features and makes the model pay more attention to the feature information of fire images. Then, label smoothing based on the cross-entropy loss function is introduced into this study to avoid predicting labels too confidently in the training process to improve the generalization ability of the recognition model. The experimental results show that the fire image recognition accuracy based on the CBAM-EfficientNetV2 model reaches 98.9%. The accuracy of smoke image recognition can reach 98.5%. The accuracy of small target detection can reach 96.1%. At the same time, we compared the existing methods and found that the proposed method achieved higher accuracy, precision, recall, and F1-score. Finally, the fire image results are visualized using the Grad-CAM technique, which makes the model more effective and more intuitive in detecting tiny targets.