DRQ-RTDETR: Degradation-Aware Detail Recovery and Query-Guided RT-DETR for Household Gas Facility Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual identification of household gas facilities is crucial for safety inspections.
• Real-world indoor images present significant challenges including clutter, shadows, reflections, and motion blur.
• Existing models like RT-DETR struggle with degraded image conditions, impacting detection accuracy.

Purpose of the Study:

• To develop an enhanced object detection method for reliable visual identification of household gas facilities.
• To address the limitations of current models in handling degraded indoor environments.
• To improve the accuracy of detecting small and occluded gas facility components.

Main Methods:

• Proposed DRQ-RTDETR model integrating degradation-aware detail recovery, small-object-guided query selection, and scale-adaptive geometric refinement.
• Coordinated intervention at feature fusion, query allocation, and box regression stages.
• Experiments conducted on a dataset of 21,813 images with 47,169 instances across eight categories.

Main Results:

• DRQ-RTDETR improved mean Average Precision (mAP) from 0.6168 to 0.6576 compared to RT-DETR.
• Significant gains observed in mAP50 (0.7909 to 0.8124) and mAP75 (0.6702 to 0.7136).
• Marked improvement in detecting small objects (mAPsmall from 0.4639 to 0.5247).

Conclusions:

• The proposed coordinated approach effectively alleviates detection failures caused by degraded image conditions.
• DRQ-RTDETR demonstrates superior performance in identifying compact components and precise localization in challenging household scenes.
• The method enhances safety inspection capabilities through more robust visual identification of gas facilities.