Towards Complex Real-World Safety Factory Inspection: A High-Quality Dataset for Safety Clothing and Helmet Detection

Fusheng Yu, Xiaoping Wang, Jiang Li, Shaojin Wu, Junjie Zhang, Zhigang ZengarXiv 2023 · DOI · arXiv

Abstract

Automated safety inspection in industrial environments is critical for preventing accidents. However, existing computer vision solutions are often limited by datasets that lack diversity in protective equipment and fail to represent the visual complexity of real-world factories. This paper introduces SFCHD, a large-scale dataset for the concurrent detection of safety helmets and protective clothing. Additionally, we propose SCALE, an attention-based module designed to improve object detection performance in the low-light conditions common in industrial settings.

Introduction

In high-risk environments such as chemical plants, ensuring that personnel wear proper Personal Protective Equipment (PPE) is a primary safety requirement. While deep learning has enabled significant progress in object detection, its deployment in industrial safety monitoring faces two major hurdles. First, existing datasets primarily focus on helmets, neglecting safety clothing. Second, most benchmarks utilize "clean" images from the web, which do not account for the noise, poor lighting, and complex backgrounds found in actual surveillance footage.

The SFCHD Dataset: Bridging the Reality Gap

To address these limitations, we present the Safety Factory Clothing and Helmet Dataset (SFCHD). This dataset comprises 12,373 images with over 50,000 annotations across seven categories. Unlike previous benchmarks, SFCHD is sourced directly from 40 different surveillance scenes within two authentic chemical plants.

A complex surveillance scene from a real-world industrial environment, illustrating the visual noise and scale challenges addressed by SFCHD.

A key feature of SFCHD is the inclusion of labels for Safety Clothing. Furthermore, we introduced specific categories for Blurred Clothing and Blurred Heads. This strategy allows models to learn features from instances obscured by shadows or motion, rather than treating them as background noise.

The SCALE Module: Enhancing Low-Light Detection

Industrial surveillance often suffers from extreme lighting conditions, including deep shadows and overexposure. To mitigate this, we developed the Spatial and Channel Attention-based Low-light Enhancement (SCALE) module. SCALE is a "plug-and-play" component that can be integrated into standard detectors like YOLO or Faster R-CNN.

The module operates through two parallel pathways:

  1. Spatial Attention Pathway (SAP): Utilizes multi-scale convolutional kernels to focus the model on salient regions, effectively ignoring dark or noisy backgrounds.
  2. Channel Attention Pathway (CAP): Learns to prioritize informative feature channels—such as specific colors or textures—to help the model "see" through low-visibility conditions.

Notably, SCALE is trained using standard detection loss, requiring no specialized loss functions or separate image-brightening pre-processing steps that might otherwise amplify noise.

Experimental Evaluation

Our experiments demonstrate that SFCHD is a significantly more challenging benchmark than existing datasets. State-of-the-art models like TOOD and VFNet achieved mean Average Precision (mAP) scores of approximately 0.51–0.52 on SFCHD, which is considerably lower than their performance on simpler, web-mined datasets.

Workers wearing safety helmets and high-visibility clothing in a challenging outdoor industrial setting, representing the target detection scenarios in the SFCHD dataset.

The SCALE module provided measurable improvements across various architectures. When added to YOLOv3 on the ExDark low-light dataset, mAP increased from 76.4% to 79.0%. On the SFCHD dataset itself, SCALE improved YOLOv8 performance by 1.1% mAP, proving its effectiveness even in modern, optimized models.

Conclusion

The introduction of SFCHD and the SCALE module represents a significant step toward robust, real-world industrial safety monitoring. By providing a high-quality, realistic dataset and a specialized enhancement module, we enable AI systems to better handle the complexities of factory environments. Future work will focus on expanding the dataset to include instance segmentation and developing methods to handle a broader range of environmental noise.

Summary updated

Abstract

Detecting safety clothing and helmets is paramount for ensuring the safety of construction workers. However, the development of deep learning models in this domain has been impeded by the scarcity of high-quality datasets. In this study, we construct a large, complex, and realistic safety clothing and helmet detection (SFCHD) dataset. SFCHD is derived from two authentic chemical plants, comprising 12,373 images, 7 categories, and 50,552 annotations. We partition the SFCHD dataset into training and testing sets with a ratio of 4:1 and validate its utility by applying several classic object detection algorithms. Furthermore, drawing inspiration from spatial and channel attention mechanisms, we design a spatial and channel attention-based low-light enhancement (SCALE) module. SCALE is a plug-and-play component with a high degree of flexibility. Extensive evaluations of the SCALE module on both the ExDark and SFCHD datasets have empirically demonstrated its efficacy in enhancing the performance of detectors under low-light conditions. The dataset and code are publicly available at https://github.com/lijfrank-open/SFCHD-SCALE.
BibTeX
@misc{yu2023towards,
  author = {Fusheng Yu and Xiaoping Wang and Jiang Li and Shaojin Wu and Junjie Zhang and Zhigang Zeng},
  title = {{Towards Complex Real-World Safety Factory Inspection: A High-Quality Dataset for Safety Clothing and Helmet Detection}},
  year = {2023},
  eprint = {2306.02098},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.02098},
}
Figure