Steel Strip Defect Sample Generation Method Based on Fusible Feature GAN Model under Few Samples.

Area of Science:

• Metallurgical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Defect identification in hot-rolled strip production suffers from limited diverse sample data due to high labeling costs.
• Insufficient data diversity negatively impacts the accuracy of steel surface defect classification models.

Purpose of the Study:

• To address the challenge of insufficient defect sample data for steel surface defect identification and classification.
• To propose a novel generative adversarial network (GAN) based model for generating diverse and high-quality defect images.

Main Methods:

• Developed the Strip Steel Surface Defect-ConSinGAN (SDE-ConSinGAN) model, a single-image GAN trained with image-feature cutting and splicing.
• Incorporated a size-adjustment function and channel attention mechanism to enhance defect feature representation.
• Dynamically adjusted training iterations to optimize training time.

Main Results:

• SDE-ConSinGAN effectively enriches image datasets with generated defect samples.
• The generated defect images exhibit superior quality and diversity compared to existing methods.
• The synthesized samples are suitable for training deep learning-based automatic classification systems.

Conclusions:

• The SDE-ConSinGAN model offers a viable solution for augmenting limited datasets in steel surface defect detection.
• Improved data diversity through SDE-ConSinGAN leads to enhanced accuracy in defect classification.
• This method holds significant potential for advancing automated quality control in the metallurgical industry.