This paper proposes a real-time fire recognition framework designed for surveillance systems by integrating DenseNet and Transformer architectures. DenseNet-121 is used to extract spatial features from video frames, while a Transformer encoder captures temporal dependencies across frame sequences. The model is trained and evaluated on the ONFIRE 2025 dataset, which provides temporally labeled surveillance videos depicting diverse fire and smoke conditions. Experimental results show that the proposed DenseNet-Transformer model achieves high recall and reasonable F1-score in fire detection tasks, demonstrating the potential of attention-based architectures in real-time applications. The study also identifies key challenges such as false alarms and limited training data, and proposes future enhancements including advanced labeling, loss functions, and post-processing strategies for improved precision and deployment readiness.