A Post-Training Channel Pruning Method Based on Grad-CAM and Its Application in Fire Detection

This paper proposes a post-training structured channel pruning scheme leveraging Gradient-Weighted Class Activation Mapping (Grad-CAM) for few-class fire detection under limited computational resources. After standard training, category-specific gradient signals extracted from the detection heads are used to generate channel-wise class activation maps over multi-scale feature layers, and fire and smoke responses are fused to rank channel importance. A layer-level retention quota is further applied to implement structured channel pruning, followed by lightweight fine-tuning. The pipeline does not require additional sparsity-inducing training. We validate the method on a self-established fire and smoke dataset containing 9041 images, using YOLOv5s, YOLOv5m, and YOLOv5l as baseline detectors. In workflow-level comparisons, the proposed method achieved higher [email protected] than the implemented L1-based workflow at 40% and 60% pruning, but not at 80%. At 60% pruning, the pruned YOLOv5s model contained 2.158 M parameters and required 2.901 GFLOPs. These results indicate that Grad-CAM provides a useful class-aware criterion for channel importance and offers a promising model-compression strategy for resource-constrained fire detection, although physical edge-device performance remains to be evaluated.