Real Time Fire and Smoke Detection on RGB Images Using YOLOv7 with Optimized Hyper-Parameters

In this study, various experiments were carried out for fire and smoke detection by using YOLOv5 and YOLOv7 pre-trained weights. As a result of the experiments, the performances of the models were evaluated separately and compared with each other. A total of 7 weights, including small, medium and large, were used for pre-trained models. In the experiments with YOLOv5, the model with the best mAP.5 score was yolov5m with 0.79, while the model with the best mAP.5 score in the YOLOv7 models was yolov7 with 0.78. Since the YOLOv7 model performed worse than expected, a hyper-parameter search was performed for yolov7 using Genetic Algorithm (GA). After hyper-parameter search with yolov7, the precision and mAP.5 value of the model increased by 0.01, while the recall value increased by 0.02. It has been observed that tuned hyper-parameters strong claim the default hyper-parameters for fire-smoke detection.