Research on Grayscale Images Flame Detections Based on Colorazion Algorithm and YOLOv8

Flame detection algorithms are crucial for real-time fire monitoring using surveillance cameras. Current flame detection algorithms perform excellently on color cameras; however, many night vision cameras can only capture grayscale images, making flame detection less effective in such scenarios. Traditional flame detection algorithms rely heavily on color features, especially the red and yellow hues of flames. Conversely, grayscale images lack color information, making these algorithms ineffective. To address this limitation, we propose a novel flame detection algorithm that integrates the DeOldify model with the you only look once(YOLOv8) model. DeOldify can effectively add color to grayscale images. YOLOv8 is a powerful object detection model that excels in identifying various objects within an image. This algorithm leverages DeOldify's colorization capabilities and YOLOv8's powerful object detection abilities to enhance flame detection accuracy in grayscale images. This algorithm has been tested multiple times on public datasets, converting original images to grayscale and then colorizing them to construct the dataset. The results show that, compared to grayscale images, this algorithm improves the mean average precision ([email protected]) by 19.7% on different validation sets, by 4% on different training sets, and by 12.9% on grayscale images captured by night vision cameras. By comparing the training results of models such as YOLOv8n and YOLOv8s, the colorized images improve the model's performance. This demonstrates the high generalizability of this algorithm.