Hardhat10K: A Large-scale Dataset for Deep Learning-based Hardhat-wearing Detection

Hardhats are crucial personal protective equipment in various industries, including construction and manufacturing. Utilizing deep learning methods to develop unattended detection systems can ensure compliance with hardhat-wearing. However, a deep-learning-based hardhat-wearing detection model requires enormous training images. Unfortunately, existing public datasets lack sufficient images and labels to train a hardhat-wearing detection model with good robustness. The introduction of the Hardhat10K dataset, comprising over 10,000 images, is a significant feature of this research. Images containing long-distance, occluded, dense, and low-light objects were collected to enhance the model's robustness. Furthermore, images from various weather conditions and periods were added to improve the model's generalization ability. Finally, background images were supplemented to enhance the model's accuracy. Compared to the state-of-the-art SHEL5K dataset, the number of images and labels increased by over one time and approximately 69.2%, respectively. The proposed dataset was evaluated using three types of models. Each model achieved good accuracy and usability on the proposed dataset. The dataset is publicly available at https://github.com/bobo504/hardhat10k.