LHO-net: A Lightweight Steel Defect Detection Framework Based on Cross-Scale Feature Selection and Adaptive...

Area of Science:

• Computer Vision
• Materials Science
• Artificial Intelligence

Background:

• Existing steel surface defect detection models struggle with complex scenarios, high computational demands, and deployment on terminal devices.
• There is a need for accurate and efficient detection methods suitable for industrial applications with limited resources.

Purpose of the Study:

• To propose a novel lightweight detection network, LHO-net, for steel surface defect detection.
• To enhance adaptability to complex scenarios, reduce computational complexity, and facilitate terminal deployment.

Main Methods:

• Developed LHO-net incorporating a Lightweight Multi-Backbone (LM Backbone), Hierarchical Scale-based Pyramid Attention Network (HSPAN), and Occlusion-aware Detection Head (OAHead).
• LM Backbone uses a dual-branch structure with dynamic feature fusion for multi-dimensional defect feature capture and parameter compression.
• HSPAN employs dynamic feature selection and adaptive upsampling for efficient multi-scale feature fusion.
• OAHead utilizes deep feature aggregation and exponential normalization for adaptive compensation of occluded defect features.

Main Results:

• LHO-net achieved mAP@0.5 of 75.0% and mAP@0.5:0.95 of 44.0% on the NEU-DET dataset with only 2.3 GFLOPS.
• Reduced parameters by 64% and computational cost by 60.3% compared to YOLOv12.
• On the GC-10 dataset, achieved mAP@0.5 of 67.2% and demonstrated superior detection stability for complex defects like slender creases and low-contrast water spots.

Conclusions:

• LHO-net effectively addresses the trade-off between detection accuracy and lightweight deployment for industrial steel surface defect detection.
• Provides an efficient and practical solution for real-time defect detection on resource-constrained terminal devices.
• Demonstrates robust performance, avoiding issues like redundant annotations and false detections.