PPFS-YOLO: Physics-Prior Frequency-Spatial Fusion for Robust Container Surface Damage Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Structural Health Monitoring

Background:

• Container surface damage detection is vital for intermodal freight transport safety.
• Existing methods struggle with pseudo-textures (false positives) and scarce defect types (false negatives).
• YOLO-family detectors lack frequency-domain analysis and physical prior integration.

Purpose of the Study:

• To develop an advanced object detection framework for container surface damage.
• To address challenges posed by visual pseudo-textures and underrepresented defect classes.
• To improve the accuracy and reliability of automated damage assessment systems.

Main Methods:

• Proposed PPFS-YOLO, a physics-prior frequency-spatial fusion framework based on YOLOv12s.
• Introduced Frequency-Spatial Fusion (FSF) module to suppress pseudo-textures using spectral masks.
• Integrated Edge-Guided Auxiliary Supervision Module (FIM) with physics-informed regularization (Lphy) for edge boundary accuracy.

Main Results:

• PPFS-YOLO achieved 64.86% mAP@50, a +12.35 pp improvement over the YOLOv12s baseline.
• The model showed a +12.10 pp gain attributed to the Lphy regularization, highlighting its critical role.
• Achieved significant performance gains with minimal parameter increase (+0.79 M) and modest latency overhead.

Conclusions:

• PPFS-YOLO effectively detects container surface damage by synergistically combining frequency-domain analysis and physics-based priors.
• The proposed FSF and FIM modules significantly enhance detection accuracy, particularly for challenging defect types.
• This framework offers a robust and efficient solution for automated structural integrity assessment in freight transport.