A novel algorithm [high-performance graphical processing unit net version 2-efficient multi-branch and scale feature pyramid network-multipath coordinate attention-you only look once (HGNetV2-EMBSFPN-MPCA-YOLO, HEM-YOLO)] was used in this study to improve smoke and fire detection accuracy in complex environments. By modifying the weights of negative and positive samples, we identified small target classes using improved exponential moving averages with spatial learning loss (EMASlide loss). Multipath coordinate attention (MPCA) was also employed to improve detection accuracy as it efficiently extracted local and global features from images. The backbone network in HGNetV2 enabled more efficient and rapid training of the algorithm. As a result, the enhanced detector head model in EMBSFPN recognized complex patterns to detect and identify smoke and fires in complex environments. The enhanced HEM-YOLO reduced the number of parameters to 1.8 million and floating point operations per second to 6.0 G and increased the accuracy by 4.2% in smoke and fire detection. Its efficiency was further improved, reducing false detections and demonstrating its versatility across multiple applications.