Efficient and accurate semi-supervised semantic segmentation for industrial surface defects.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Deep learning methods are crucial for industrial quality inspection but face challenges with limited data, low utilization, and performance bottlenecks.
• Existing semi-supervised methods struggle with accuracy and speed in complex industrial defect detection scenarios.

Purpose of the Study:

• To propose a novel semi-supervised semantic segmentation framework to address limitations in industrial defect detection.
• To improve accuracy, data utilization, and real-time performance in automated quality inspection.

Main Methods:

• Developed a semi-supervised semantic segmentation framework utilizing perturbation invariance in image and feature spaces.
• Employed diverse perturbation cross-pseudo-supervision to minimize reliance on large labeled datasets.
• Integrated edge pixel-level semantic information and shallow feature fusion for enhanced efficiency and accuracy.

Main Results:

• The proposed method significantly outperforms state-of-the-art (SOTA) semi-supervised semantic segmentation techniques in industrial settings.
• Achieved a mean Intersection over Union (mIoU) of 3.11% higher than SOTA on a custom dataset and 4.39% higher on the KolektorSDD dataset.
• The network matches U-net's speed while surpassing DeepLabv3Plus in mIoU by 2.99%.

Conclusions:

• The developed framework offers a robust and efficient solution for industrial defect detection using semi-supervised learning.
• The method effectively handles insufficient sample sizes and improves detection of small targets and defect edges.
• This approach advances automated quality inspection by enhancing both accuracy and real-time processing capabilities.