WCSE

Identification of Smoke and Fire from a Low-Light Enhanced Synthetic Dataset Using YOLOv7

Abstract— This paper investigates the utilization of YOLOv7, a state-of-the-art object detection system, in identifying instances of smoke and fire using a synthetic low-light enhanced dataset. With fire and smoke incidents causing significant casualties annually, the study aims to enhance the efficacy of fire detection systems. Through extensive experimentation and analysis, the study compares the performance of YOLOv7 on original and low-light enhanced datasets. Results indicate that while the model shows promise, the lowlight enhancements, particularly Contrast Limited Adaptive Histogram Equalization (CLAHE), outperform deep learning-based enhancements. Further research avenues include exploring more sophisticated low-light image enhancement techniques, employing diverse and balanced datasets, and deploying the model in realworld settings for comprehensive evaluation and practical applications. This study contributes to advancing non-urban fire detection systems, thereby bolstering public safety measures and disaster management strategies.

Index Terms— Fire Detection, Smoke Detection, Wildfire Detection, Computer Vision, YOLOv7, Vision Sensor-Based Techniques, Low-Light Conditions, Synthetic Dataset, Object Detection.

John Paul Q. Tomas
Mapua University, PHILIPPINES
Von Derwin R. Caibigan
Mapua University, PHILIPPINES
Erin Beatrice Micaela G. Reyes
Mapua University, PHILIPPINES

[Download]

Cite: John Paul Q. Tomas, Von Derwin R. Caibigan, Erin Beatrice Micaela G. Reyes, "Identification of Smoke and Fire from a Low-Light Enhanced Synthetic Dataset Using YOLOv7," 2024 The 14th International Workshop on Computer Science and Engineering (WCSE 2024), pp. 50-57, Phuket Island, Thailand, June 19-21, 2024.