An open-access fire detection image dataset for research in computer vision and safety monitoring

This article presents the Fire Recognition Image Dataset, a comprehensive collection of images designed to support research in fire detection, smoke recognition, and safety monitoring using computer vision techniques. The dataset was created by systematically selecting publicly available YouTube videos depicting both fire and non-fire scenarios. Sources included uncontrolled fires such as forest, industrial, and household incidents, as well as controlled fire scenes such as campfires, stoves, and candle flames. Frames were manually extracted at regular intervals using Python's OpenCV library, with redundant, low-quality, or irrelevant frames removed to ensure high-quality data. The dataset comprises 1112 original RGB images categorized into four classes: Real Fire, Smoke, Safe Fire, and Artificial Fire, covering a wide range of environmental conditions, fire types, and perspectives. To increase variability and improve robustness for machine learning applications, a structured data augmentation pipeline was applied using the Albumentations library, including horizontal and vertical flips, rotations, affine shear transformations, brightness and contrast adjustments, and Gaussian noise addition. This augmented the dataset to 6672 additional images, resulting in a total of 7784 images. All images are standardized to 400 × 400 pixels and stored in PNG format. The dataset is organized into class-specific folders with separate subfolders for original and augmented images, enabling straightforward integration into training pipelines. Supporting scripts for frame extraction, preprocessing, and augmentation are provided to ensure reproducibility. This openly accessible dataset offers a valuable resource for researchers developing and benchmarking deep learning models for binary and multiclass fire detection tasks. It can be reused for applications in environmental monitoring, disaster management, and intelligent surveillance, and serves as a benchmark for evaluating computer vision algorithms under diverse real-world and controlled fire conditions.