ZoomHead: A Flexible and Lightweight Detection Head Structure Design for Slender Cracks

Area of Science:

• Computer Vision
• Deep Learning
• Materials Science

Background:

• Metal surface crack detection is crucial for industrial equipment safety.
• Existing deep object detection models are often too complex and resource-intensive for real-world deployment.
• There is a need for efficient and accurate defect detection solutions.

Purpose of the Study:

• To optimize the YOLO series head structure for improved efficiency and performance.
• To develop a lightweight detection head, ZoomHead, with reduced computational complexity and enhanced detection capabilities.
• To enable practical deployment of advanced defect detection on frontline construction sites.

Main Methods:

• Replaced BatchNorm2d with GroupNorm2d for stable feature distribution and faster training.
• Introduced Detail Enhanced Convolution (DEConv) and shared convolution to improve detail capture and reduce redundancy.
• Integrated a Zoom scale factor in the regression branch to minimize computational complexity.
• Applied ZoomHead to YOLOv10 and YOLOv11 models for comparative experiments on rail and NEU surface defect datasets.

Main Results:

• ZoomHead integration improved detection accuracy and Frames Per Second (FPS).
• The proposed structure reduced model parameters and computational load.
• Achieved a balance between detection accuracy and speed, outperforming State-of-the-Art (SOTA) models in efficiency.
• Maintained high mean Average Precision (mAP) with fewer parameters and higher FPS compared to SOTA.

Conclusions:

• ZoomHead offers superior comprehensive detection performance for metal surface defects.
• The lightweight design makes it suitable for deployment in resource-constrained environments.
• This optimization addresses the limitations of current deep learning models in industrial defect detection.