Constructing a Lightweight Fire and Smoke Detection Through the Improved GhostNet Architecture and Attention Module Mechanism

Fire disasters have a serious impact on human life and property. Fire detection methods have been developed as early detection in preventing fire disasters, including implementing object detection models. However, the challenges in fire detection are random object shapes and complex backgrounds, which cause the model to be unable to detect fire optimally. In addition, to support practical application, the model must be efficient in memory-consuming. To address these issues, this study proposed FSP-YOLO as a fire detection algorithm that integrates the YOLOv8 model with the improved GhostNet (iGhostNet) and the attention module that can detect three classes, including fire, smoke, and people. The improvement of iGhostNet is implementing coordinate attention module and trained integrated learning, which has an impact on increasing accuracy and reducing computational costs. With this improvement, the proposed fire detection model could gain outstanding performance, reducing computational costs, and a lightweight architecture. Furthermore, to validate the performance of FSP-YOLO, several comparative experiments have been conducted, including single and multiple module experiments, evaluation under different YOLOv8, experiments with other object detection algorithms, and experiments in practical application. From these experiments, FSP-YOLO represents the optimal performance, which results in mAP0.5 of 0.977, precision of 0.952, and recall of 0.973. Based on weight or model size, compared to the other fire detection algorithms such as YOLOFM, FireRPG, Fine Tuned YOLOv8, Improved YOLOv8, YOLOv7-Fire, YOLOv9, and Improved-YOLOv7, FSP-YOLO has a lighter size of 53.18 MB. Therefore, FSP-YOLO is reliable in accuracy but reduces the computational cost.