A Novel ST-YOLO Network for Steel-Surface-Defect Detection

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Deep learning shows promise in defect detection but faces challenges with complex image features and dynamic model changes.
• Existing methods struggle with the rich semantic levels and diverse morphological characteristics of defect images.

Purpose of the Study:

• To propose an effective deep learning model, ST-YOLO, for enhanced steel defect detection.
• To address challenges in defect detection, including complex image features and dynamic model learning.

Main Methods:

• Introduced a shunt feature fusion model (ST-YOLO) with a split feature network structure.
• Implemented a self-correcting transmission allocation method for training, including adaptive sampling and dynamic label allocation.
• Designed the network to specialize classification and localization tasks based on computational requirements.

Main Results:

• The ST-YOLO model demonstrated superior performance on the NEU-DET and GC10-DET datasets.
• The proposed adaptive training strategy effectively utilized high-quality samples to optimize the training process.
• The model's specialized network structure improved the handling of diverse defect features.

Conclusions:

• ST-YOLO offers an effective solution for steel defect detection, outperforming existing methods.
• The adaptive sampling and dynamic label allocation contribute to robust and accurate defect identification.
• The model's architecture is well-suited for handling the complexities of defect image analysis.