Few-Shot Strip Steel Surface Defect Segmentation via Pre-Trained Variational Auto-Encoder-Based Latent Gaussian...

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Few-shot segmentation for strip steel surface defects is crucial but challenging.
• Existing methods often rely on frozen encoders pre-trained on classification, limiting knowledge transfer.
• Effective characterization of defect regions requires richer image-related knowledge.

Purpose of the Study:

• To propose a novel method for few-shot strip steel surface defect segmentation.
• To leverage a pre-trained Variational Auto-Encoder (VAE) for enhanced feature representation.
• To introduce latent Gaussian process regression (LGPR) for efficient feature correlation.

Main Methods:

• Utilized a frozen VAE (encoder and decoder) pre-trained with a pixel-level self-supervised image reconstruction task.
• Employed Gaussian process regression in the VAE's latent feature space to build correlations between support and query features.
• Integrated transformer-based projectors to capture long-range contextual information.

Main Results:

• The proposed LGPR method significantly outperforms state-of-the-art models on two public datasets.
• The VAE-based encoder and decoder provide rich image-related knowledge for defect characterization.
• Non-parametric Gaussian process regression in latent space efficiently builds pixel-level correlations without additional training overhead.

Conclusions:

• The developed LGPR method is highly effective for few-shot strip steel surface defect segmentation.
• Pre-training with self-supervised tasks and utilizing latent Gaussian process regression are key to the model's success.
• The approach offers a robust and efficient solution for industrial surface inspection.