Semi-Supervised Learning for Defect Segmentation with Autoencoder Auxiliary Module.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Unsupervised anomaly detection methods often struggle with rare defect samples in datasets.
• Traditional autoencoder approaches for anomaly detection may not fully utilize available ground truth data.
• Determining appropriate thresholds for anomaly segmentation can be challenging.

Purpose of the Study:

• To propose a semi-supervised learning framework for automatic defect region segmentation.
• To effectively utilize both unlabeled and limited labeled datasets for improved anomaly detection.
• To enhance the accuracy and stability of defect segmentation in images.

Main Methods:

• A semi-supervised approach combining an autoencoder for normal data reconstruction and a segmentation module.
• Training an autoencoder on unlabeled, predominantly normal samples to reconstruct defect-free images.
• Utilizing a difference map between input and reconstructed images, along with ground truth, to train a segmentation network with binary cross-entropy loss.

Main Results:

• The proposed semi-supervised method demonstrates improved defect segmentation compared to unsupervised approaches.
• Experimental results show an overall performance improvement of 3.83% with the aid of limited ground-truth segmentation maps.
• The integration of difference images enhances training stability.

Conclusions:

• Semi-supervised learning offers a powerful approach to anomaly detection by leveraging both unlabeled and labeled data.
• The proposed method effectively segments defect regions, outperforming traditional unsupervised techniques.
• Even a small amount of labeled data can significantly boost the performance of anomaly detection systems.