A lightweight defect detection algorithm for escalator steps.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Mechanical Engineering

Background:

• Escalator step defect detection faces challenges with large model parameters, poor adaptability, and real-time video processing.
• Existing deep learning models struggle with the small size and subtle nature of escalator step defects.

Purpose of the Study:

• To propose an efficient target detection algorithm, ASF-Sim-YOLO, for real-time escalator step defect detection.
• To enhance detection accuracy for small targets and improve model efficiency for mobile deployment.

Main Methods:

• Designed the ASF-Sim-P2 structure to improve small target detection.
• Integrated Similarity-based Attention Mechanism (SimAM) with Spatial Pyramid Pooling-Fast (SPPF) for enhanced feature capture.
• Replaced Complete-Intersection-over-Union (CIoU) loss with Normalized Wasserstein Distance (NWD) to reduce missed defects.
• Applied channel pruning for lightweight model design suitable for mobile devices.

Main Results:

• Achieved 96.8% mAP50 accuracy, a 22.1% improvement over the baseline.
• Reduced computational complexity to one-quarter of the baseline.
• Increased frame rate (FPS) to 575.1.
• Outperformed YOLOv3-tiny, YOLOv5s, YOLOv8s, Faster-RCNN, TOOD, and RTMDET in accuracy and real-time capability.

Conclusions:

• ASF-Sim-YOLO effectively balances lightweight design with performance improvements for escalator step defect detection.
• The algorithm is highly suitable for real-time inspection operations, meeting the demands of escalator maintenance.