Hierarchical depth aware YOLO for efficient metal surface defect detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Metal surface defect detection is crucial for intelligent manufacturing.
• Existing lightweight detectors struggle with complex backgrounds and variations.
• Uniform feature refinement limits performance and efficiency.

Purpose of the Study:

• To develop an effective lightweight metal surface defect detection framework.
• To address the limitations of uniform feature refinement in existing detectors.
• To improve the balance between fine-grained representation and computational efficiency.

Main Methods:

• Proposed a hierarchical depth-aware refinement framework (HDR-YOLO).
• Introduced Query-Focused Convolution (QFC) for shallow layers (texture/edge enhancement).
• Employed Query-Based Fusion (QBF) for deep layers (semantic modeling).

Main Results:

• HDR-YOLO significantly improved detection performance on NEU-DET and GC10-DET datasets.
• Achieved mAP@0.5 improvements of 3.92% and 7.67% over the baseline.
• Maintained competitive inference efficiency, suitable for real-time industrial constraints.

Conclusions:

• Depth-aware refinement is an effective strategy for lightweight defect detection.
• HDR-YOLO enhances detection of small-scale and irregular defects.
• The proposed framework offers a practical solution for real-time industrial defect detection.