GPAC-YOLOv8: Lightweight Target detection for fire scenarios

Abstract Current fire object detection methods face challenges due to the large number of parameters in deep network models, making it difficult to adapt to limited hardware configurations. Additionally, detecting small targets in the early stages of a fire is challenging. Therefore, this paper proposes a fire and smoke detection method based on
YOLOv8, named GPAC-YOLOv8. Firstly, the Ghost module and PSA attention mechanism are integrated into the Backbone to design the CGP module, which enhances computational speed without sacrificing accuracy. Then, a new feature fusion module, AC-Neck, is designed using the ASFF method and the lightweight CARAFE upsampling module to optimize feature map fusion and improve the performance of small target detection. Finally, a Focal-WIoU loss function with a dual weighting mechanism is designed to accurately define the aspect ratios of the predicted bounding boxes, thereby enhancing the model's generalization capability . Experimental results using the proposed
GEAC-YOLOv8 method on a self-made dataset show significant improvements in detection speed while maintaining detection accuracy compared to other t raditional methods. Therefore, the GPAC-YOLOv8 method effectively enhances the performance of object detection in fire and smoke scenarios.