SH-DETR: Enhancing steel surface defect detection and classification with an improved transformer architecture

Area of Science:

• Computer Vision
• Deep Learning
• Materials Science

Background:

• Steel surface defect detection is crucial but challenging due to defect complexity and variety.
• Conventional models struggle with low recognition accuracy and insufficient classification power.
• Existing methods require advanced techniques for accurate defect identification.

Purpose of the Study:

• To develop a deep learning framework for enhanced steel surface defect detection.
• To address limitations of conventional models in accuracy and classification power.
• To improve the efficiency and effectiveness of identifying diverse steel surface defects.

Main Methods:

• Proposed a novel deep learning framework combining multi-channel random coding and multi-scale feature fusion.
• Integrated Transformer architecture's self-attention mechanism with Convolutional Neural Networks (CNNs) using ResNet18.
• Introduced a multi-channel shuffled encoding module and an upsample concatenated Simple Parameter-Free Attention Module (UPC-SimAM) for feature extraction and fusion.

Main Results:

• The SH-DETR model demonstrated superior performance on NEU-DET and GC10-DE datasets compared to state-of-the-art methods.
• Achieved 91.72% classification accuracy, 83.03% mAP@0.5, and 45.55% mAP@0.5:0.95 on the NEU-DET dataset.
• Obtained 76.73% classification precision, 65.03% mAP@0.5, and 32.46% mAP@0.5:0.95 on the GC10-DE dataset.

Conclusions:

• The proposed SH-DETR model significantly enhances steel surface defect detection efficiency and accuracy.
• Ablation studies and visualization confirmed the model's effectiveness and potential.
• The framework offers a promising solution for complex defect identification in computer vision.