Abstract: Abstract: To address the issues of flame and smoke occlusion， difficulty distinguishing between flame/smoke and background， and low detection accuracy in complex fire scenarios， a detection model called CFD-YOLO（Complex Fire Detection YOLO）has been proposed. This model is based on the YOLOv8 framework， with several enhancements. Firstly， in the backbone network， the model combines Deformable Convolution DCNv4 with the C2f module， leveraging DCNv4's unconstrained dynamic weighting mechanism to significantly improve the capture of complex deformations and non-rigid features in fire images. Secondly， in the neck section， a cross-attention-based deep semantic fusion module， PSFM， is introduced to achieve adaptive feature enhancement by deeply fusing different feature layers of fire images. Finally， in the head section， the occlusion-aware attention SEAM is used to improve the detection head， allowing it to effectively handle occlusions of flames and smoke in complex environments. The loss function employed is SlideLoss， which dynamically adjusts the positive and negative sample coefficients to reduce false detection rates. The experimental results showed that the mAP index reached 80.33% and 88.28% respectively in the self built dataset and public dataset， which were 3.85 and 3.91 percentage points higher than the original YOLOv8 network. Compared with the current mainstream models， it also has good detection performance.

Key words: Key words： fire detection, deep semantic fusion module, attention mechanism, deformable convolution, YOLOv8