HIFE-YOLO: HierarchicalInformation Flow Enhancement for Lightweight Fire and Smoke Detection

Fire and smoke detection requires compact models that preserve weak local evidence while using sufficient context to distinguish diffuse targets. This paper proposes HIFE-YOLO, a YOLO11n-based detector that coordinates multiscale Ghost feature generation, retention-path-assisted neck downsampling, and complementary task residuals before the detection towers. The fused model contains 2.303 million parameters and has an estimated computational cost of 6.062 GFLOPs for a 640×640 input, representing reductions of 10.84% and 4.00%, respectively, relative to YOLO11n. Under a common D-Fire training protocol of 100 epochs, three paired training seeds yield an mAP50 of 73.25±0.36% and an mAP50–95 of 41.38±0.44%, compared with 71.95±0.23% and 40.16±0.19% for the baseline. Fire and smoke AP50 for relatively small objects improve by 1.41 and 4.49 percentage points, respectively. On the independent fire_merged held-out view, frozen-weight testing improves overall mAP50 and recall by 1.74 and 5.67 percentage points, respectively, while precision and the HPWREN smoke AP do not improve. The parameter and storage savings are accompanied by slightly higher latency and peak GPU allocation in the measured desktop pipeline. The relatively small-object tests and the BoWFire fire-presence test show clear gains, and all conclusions remain limited to the D-Fire training protocol and the recorded seeds.