An Efficient Fire Detection Algorithm with Dataset Generation Based on Diffusion

Faced with increasing urbanization and population growth, the hazard of fires has become a critical issue in urban areas worldwide. Therefore, the focus on how to achieve accurate and rapid fire detection using artificial intelligence has intensified. Currently, fire detection algorithms often struggle to balance precision and processing speed. At the same time, high-quality fire detection datasets are extremely scarce, severely limiting the performance of algorithms. To address these challenges, this paper first proposes an enhanced fire detection algorithm framework based on YOLOv8, aiming to comprehensively improve the accuracy and processing speed of detection. This framework integrates Swin Transformer and Slim Neck to enhance model performance, and utilizes WIoUv3 for network training regularization. Additionally, this paper introduces a novel fire detection dataset generation algorithm: Fire-Generator, to address the scarcity of high-quality training data. This tool utilizes a latent diffusion model to guide image generation, enabling the synthesis of training data from textual descriptions or existing images. Experimental evaluations demonstrate that our method outperforms most currently available fire detection algorithms in terms of performance. The versatility of Fire-Generator extends beyond fire detection, demonstrating potential for generating training data across different domains. Through this comprehensive strategy, our model achieves significant results by leveraging both real and synthesized data.