Triple-attentions based salient object detector for strip steel surface defects

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Accurate detection of surface defects in strip steel is critical for quality control.
• Current deep learning detectors refine features but can be further improved.
• Attention mechanisms are key for feature extraction and fusion in defect detection.

Purpose of the Study:

• To introduce an innovative triple-attention mechanism (TA) for enhanced feature representation in strip steel defect detection.
• To propose a novel detector, TADet, based on the TA mechanism.
• To improve the accuracy and robustness of strip steel surface defect detection.

Main Methods:

• Developed a triple-attention (TA) mechanism analyzing feature maps from channel-width, channel-height, and width-height perspectives.
• Proposed TADet, an encoder-decoder network incorporating TA to refine and fuse multi-scale features.
• Conducted extensive experiments to evaluate TADet's performance.

Main Results:

• TADet demonstrated superior performance over state-of-the-art methods.
• The model achieved improvements in mean absolute error, S-measure, E-measure, and F-measure.
• Experimental results confirmed the effectiveness and robustness of TADet.

Conclusions:

• The proposed triple-attention mechanism significantly enhances feature representation for defect detection.
• TADet offers a robust and effective solution for strip steel surface defect detection.
• The findings contribute to advancing automated quality inspection in steel manufacturing.