TL;DR
This paper introduces DST-DETR, a novel framework that combines image dehazing with object detection to improve safety helmet detection in foggy conditions. By enhancing image quality before detection, DST-DETR significantly boosts accuracy, especially for small and occluded helmets, making it more robust for real-world construction safety monitoring.
What Problem Does This Paper Address?
Detecting safety helmets in construction environments is crucial for worker safety. However, foggy weather severely degrades image quality, reducing visibility and making it difficult for object detection models to accurately identify helmets. This challenge is amplified because safety helmets often appear as small objects, are prone to occlusion, and can be easily confused with complex backgrounds. Existing methods often struggle with real-time performance, robustness in adverse conditions, and accurately detecting small objects in low-visibility scenarios. This paper aims to bridge this gap by developing a unified framework that addresses both image degradation and object detection challenges simultaneously.
Methods
The DST-DETR framework is composed of two main modules:
- PAOD-Net (Image Dehazing Module): This module is an enhancement of the lightweight AOD-Net.
- PfConv Module: A novel convolutional module is introduced, incorporating a parameter-free average attention module (PfAAM). This module focuses attention on crucial features without increasing model complexity, improving efficiency.
- Hybrid Loss Function (MS-SSIM + l2): To improve robustness and perceptual quality, a combination of MS-SSIM (Multi-Scale Structural Similarity) and l2 loss is used. This balances pixel-level accuracy with structural and perceptual similarity, leading to better dehazing results.
- ST-DETR (Object Detection Module): This module is an adaptation of the RT-DETR model for small-object detection.
- Variant ResNet-18 Backbone: A lighter ResNet-18 is used as the backbone to reduce computational load while maintaining accuracy.
- CCFF-BiFPN-P2 Neck: The neck structure is replaced with an improved BiFPN (Bi-directional Feature Pyramid Network) configuration. This enhances feature integration and cross-scale transfer, particularly beneficial for small objects.
- Small-Object Detection Layer: An additional detection layer is integrated to specifically improve the recognition of small objects, especially in low-resolution or degraded images.
Datasets
- Custom Dataset: The authors created a dataset of 9315 images of outdoor construction scenes. This dataset was used to build both a dehazing dataset and an object detection dataset.
- Object Detection Dataset: 7452 images were allocated for training, testing, and validation. A subset of 1490 images was used as a joint test set for both dehazing and object detection.
- Dehazing Dataset: Generated using an atmospheric scattering model with varying A (global atmospheric light) and β (scattering coefficient) values to simulate different fog densities (Light, Medium, Heavy, Multiple Haze).
- Public Datasets:
- RTTS: A real-world foggy dataset (4322 images) primarily covering traffic and driving scenes.
- SFID: A simulated foggy dataset (13,718 images) of insulator images.
Key Insights and Ideas
- Dehazing as a Preprocessing Step is Crucial: The paper strongly advocates for integrating image dehazing directly into the object detection pipeline, demonstrating that improved image quality significantly boosts detection performance, especially in challenging foggy conditions.
- Lightweight Enhancements for Real-Time Performance: The introduction of PfConv and the use of ResNet-18 as a backbone highlight a focus on achieving high performance without sacrificing computational efficiency, crucial for real-time applications.
- Tailoring Detection for Small Objects: The adaptation of RT-DETR with a BiFPN neck and a dedicated small-object layer shows the importance of specialized architectures for detecting small, often occluded, objects in complex scenes.
- Perceptual Loss for Better Restoration: The use of MS-SSIM alongside l2 loss indicates a move towards more perceptually aware image restoration, which better aligns with human visual perception and can lead to more visually pleasing and informative dehazed images.
Key Results
- Improved Dehazing Quality: The PAOD-Net module achieved superior PSNR and SSIM scores compared to traditional methods (DCP, CAP) and other deep learning models (DehazeNet, MSCNN, AOD-Net, FFA-Net), particularly in heavy haze conditions.
- Enhanced Object Detection Accuracy: DST-DETR demonstrated significant improvements in [email protected] across various haze levels compared to RT-DETR and other YOLO variants. For instance, it achieved a 4.0% increase in [email protected] under heavy haze compared to RT-DETR.
- Robustness Across Haze Densities: The framework showed strong generalization capabilities, performing well on light, medium, and heavy haze conditions, outperforming models that struggle in severe fog.
- Competitive Performance on Public Datasets: DST-DETR achieved state-of-the-art or near state-of-the-art results on the RTTS and SFID datasets, outperforming several existing methods in key metrics like Precision, Recall, and mAP.
- Real-time Capability: The lightweight design of both modules contributes to efficient processing, making the framework suitable for real-time applications.
Limitations and Future Work
- Further Improvement in Adverse Weather: While effective in fog, the authors suggest exploring ways to overcome more complex adverse weather conditions like low light, rain, and snow.
- Integration with Object Tracking: Incorporating object tracking algorithms could enhance site supervision and safety monitoring.
- Generalization to Unseen Scenarios: Applying the framework to natural foggy scenes without dedicated dehazing datasets could further improve its generalization capabilities.
- Dataset Limitations: The authors acknowledge the need for more diverse and extensive datasets for broader applicability.
Summary updated