Research on Lightweight Improvement of YOLOv8-based Fire Detection Model for Edge Computing

In the field of fire detection, computational constraints and limited memory resources pose challenges for robots and edge-embedded devices. Furthermore, some lightweight models struggle with effectively detecting complex targets. To address these issues, this study proposes a lightweight enhancement for the YOLOv8 model. The approach involves replacing the original backbone network with EfficientNetV2 for feature extraction, incorporating weighted feature fusion, and introducing the Convolutional Block Attention Module (CBAM). The improved model undergoes experimentation using a self-constructed dataset for fire detection. Results indicate a significant reduction in network parameters and model inference time compared to the original YOLOv8 model, accompanied by an increase in the processing frame rate. Moreover, the enhanced model demonstrates a 1.9% improvement in the map@50 metric and a 1.2 % enhancement in map@50-95. This affirms that the improved model not only ensures superior detection performance but is also well-suited for applications in embedded devices and edge computing scenarios.