Fire_detection_version_yolov11

This dataset is developed to support research in vision-based fire and smoke detection using deep learning models, particularly YOLO-based architectures. It contains a curated collection of images representing three classes: fire, smoke, and background. The fire class includes visible flame scenarios under diverse environmental conditions. The smoke class contains different smoke patterns, including light haze and dense smoke, with or without visible flames. The background class consists of normal scenes and hard negative samples such as sunlight reflections, clouds, fog, steam, dust, and artificial lighting, which are commonly misclassified as fire or smoke. The dataset is constructed from multiple sources to ensure diversity and robustness. A portion of the images is derived from publicly available datasets , including:- Peng, B., & Kim, T.-K. (2025). YOLO-HF: Early Detection of Home Fires Using YOLO. IEEE Access, 13, 79451–79466.- Putra, A. K. (2025). Indoor Fire Smoke Dataset. In addition, supplementary images of fire, smoke, and non-fire scenes were collected from open-access platforms, including Pixabay, Wikimedia Commons, and Pexels. All images are annotated in YOLO format using bounding boxes. Where applicable, polygon annotations were converted into bounding boxes during preprocessing to ensure compatibility with object detection frameworks. The dataset is divided into training, validation, and test sets using a 70/15/15 split. Data augmentation techniques were applied to the training set to improve generalization, including rotation, cropping, shearing, and color adjustments. The dataset was iteratively refined across multiple preprocessing versions to improve annotation quality and reduce noise. It is specifically designed to address the false positive problem in fire detection by including challenging negative samples. In addition to the image data, this repository provides annotation files and dataset splits to support reproducibility.本数据集旨在支持基于深度学习模型（尤其是基于YOLO（You Only Look Once）架构）的视觉型火灾与烟雾检测研究。数据集包含经过精心筛选的图像集，涵盖三类目标：火灾、烟雾与背景。 火灾类别包含多种环境条件下的可见火焰场景。烟雾类别涵盖不同形态的烟雾特征，包括轻度霾与浓烟雾，可伴随或不伴随可见火焰。背景类别则包含常规场景，以及易被误判为火灾或烟雾的难分负样本，例如阳光反射、云层、雾霭、蒸汽、扬尘与人工照明等。 本数据集通过多源构建以保障多样性与鲁棒性。部分图像源自公开数据集，包括： - Peng, B., & Kim, T.-K. (2025). YOLO-HF: Early Detection of Home Fires Using YOLO. IEEE Access, 13, 79451–79466. - Putra, A. K. (2025). Indoor Fire Smoke Dataset. 此外，研究团队从Pixabay、Wikimedia Commons与Pexels等开源平台收集了火灾、烟雾与非火灾场景的补充图像。 所有图像均采用YOLO格式进行边界框标注。若存在多边形标注，预处理阶段会将其转换为边界框以适配各类目标检测框架。 数据集按照70/15/15的比例划分为训练集、验证集与测试集。对训练集应用了数据增强技术以提升模型泛化能力，包括旋转、裁剪、剪切与色彩调整等操作。 本数据集历经多版预处理迭代优化，以提升标注质量并降低噪声。其专为解决火灾检测中的假阳性问题而设计，纳入了极具挑战性的负样本。 除图像数据外，本仓库还提供标注文件与数据集划分方案，以保障研究可复现性。