Feature fusion-based preprocessing for steel plate surface defect recognition.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Steel strip surface defects pose significant challenges for quality control.
• Traditional machine vision methods often struggle with complex defect features.
• Convolutional Neural Networks (CNNs) show promise but require effective preprocessing.

Purpose of the Study:

• To propose a feature fusion-based preprocessing strategy for steel strip surface defect detection.
• To enhance image feature dimensions and highlight pixel-level defect characteristics.
• To improve the recognition accuracy of CNNs in steel defect identification.

Main Methods:

• Utilized machine vision technology with feature fusion preprocessing.
• Applied image feature extraction operators (Sobel, Laplace, Prewitt, Robert, Local Binary Pattern) to extract edges and textures.
• Fused processed grayscale images with original images across three channels, then converted to a single channel using weighted ratios for computational efficiency.

Main Results:

• Explored various operator combinations and weight ratios on the NEU steel plate surface defect database.
• Identified an optimal fusion scheme (Sobel:image:Laplace) with a 0.2:0.6:0.2 weight ratio for single-channel conversion.
• Achieved a 3% improvement in recognition rate for unprocessed images, reaching a final accuracy of 99.77%.

Conclusions:

• The proposed feature fusion strategy significantly enhances steel strip surface defect detection accuracy.
• The method effectively increases feature dimensionality and highlights critical defect features for CNNs.
• The optimized fusion scheme offers a computationally efficient and highly accurate solution for industrial quality control.