Perturbed Progressive Learning for Semisupervised Defect Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Intelligent Manufacturing

Background:

• Surface defect inspection is crucial in intelligent manufacturing, with increasing demand.
• Deep learning shows promise but struggles with rare defect data and difficult pixelwise annotation.
• Existing supervised methods are often impractical due to data limitations.

Purpose of the Study:

• To propose a semisupervised defect segmentation (SSDS) method for scenarios with limited labeled data.
• To develop an efficient and simple approach named perturbed progressive learning (PPL).
• To enhance defect segmentation accuracy in industrial applications.

Main Methods:

• PPL utilizes a student-teacher network architecture to decouple predictions and reduce overfitting on noisy pseudo-labels.
• It encourages consistency across perturbations in a stagewise manner to mitigate label drift.
• The method involves two stages: initial pseudo-labeling of easy/hard unlabeled data and progressive refinement using perturbed data.

Main Results:

• PPL was evaluated on a new mobile screen defect dataset (MSDD-3) and public datasets.
• The method demonstrated significant improvements over state-of-the-art techniques across various evaluation protocols.
• Experimental results confirm the effectiveness of PPL in defect segmentation with limited labels.

Conclusions:

• Perturbed progressive learning (PPL) offers an effective solution for semisupervised defect segmentation.
• The method overcomes limitations of rare defect data and complex annotations in intelligent manufacturing.
• PPL significantly enhances defect inspection capabilities, surpassing existing approaches.