YOLOX-Ray: An Efficient Attention-Based Single-Staged Object Detector Tailored for Industrial Inspections

Area of Science:

• Computer Vision
• Artificial Intelligence
• Industrial Engineering

Background:

• Industrial inspection is vital for quality and safety.
• Deep learning models show promise for inspection tasks.
• Existing methods may struggle with multi-scale object detection in industrial settings.

Purpose of the Study:

• To introduce YOLOX-Ray, an efficient deep learning architecture for industrial inspection.
• To enhance feature extraction and small-scale object detection capabilities.
• To evaluate YOLOX-Ray's performance in real-world industrial inspection scenarios.

Main Methods:

• YOLOX-Ray architecture based on YOLO object detection.
• Integration of SimAM attention mechanism within FPN and PAN.
• Utilization of Alpha-IoU cost function for small object detection.

Main Results:

• YOLOX-Ray achieved high mAP50 scores (89% to 99.6%) in hotspot, crack, and corrosion detection.
• Significant performance on mAP50:95 metric (44.7% to 66.1%) for challenging small objects.
• Comparative analysis confirmed the synergistic benefit of SimAM and Alpha-IoU.

Conclusions:

• YOLOX-Ray demonstrates superior performance in multi-scale object detection for industrial inspection.
• The combined SimAM attention and Alpha-IoU loss are crucial for optimal results.
• YOLOX-Ray enables more effective, efficient, and sustainable industrial inspection processes.