Abstract:

To reduce the missed detection rate of fire detection in complex scenes and improve the anti-interference ability for flame and smoke targets, a fire detection algorithm based on a deep learning object detection framework was proposed. First, YOLOv8n was used as the baseline network, and a wavelet transform convolution (WTConv) module was introduced in the downsampling stage of the feature pyramid to retain high-frequency details of flame edges and smoke textures. Second, a poly-kernel lightweight Inception block (PKLBlock) was constructed in the feature fusion stage, and multi-scale feature fusion was performed through multi-branch convolution and contextual anchor attention. Third, a non-local attention (NLAttention) module was incorporated into the detection feature layer, and a dynamic attention mechanism was used to establish long-range dependencies between fire and smoke targets and surrounding regions. Finally, ablation and comparative performance evaluation were conducted using a self-built fire and smoke image dataset to evaluate the algorithm's performance. The results show that the algorithm achieves a precision of 93.46%, a recall of 88.48%, an F₁ score of 90.90%, and a mean average precision (mAP) of 92.04% at an intersection over union threshold of 0.5 (mAP@0.5). Compared with YOLOv8n, they are increased by 5.24, 2.13, 3.62 and 3.87 percentage points, respectively. The inference speed is 31.2 frames/s. The algorithm can detect flame and smoke targets in real time and provides algorithmic support for early fire warning in video monitoring scenarios.

Key words: multi-scale feature fusion, dynamic attention, fire detection, deep learning, flame and smoke detection