Patch-sampled contrastive learning for dense prediction pretraining in metallographic images.

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Microstructure characteristics are crucial for alloy mechanical properties.
• Deep learning excels in image analysis but is hindered by high annotation costs.
• Existing self-supervised learning methods need adaptation for microstructure-specific tasks.

Purpose of the Study:

• To develop a microstructure-specific pretraining framework addressing annotation costs.
• To enhance deep learning-based microstructure identification in metallographic images.
• To improve the efficiency and accuracy of microstructure segmentation.

Main Methods:

• Proposed a novel patch-sampled contrastive learning (PSCL) method.
• Implemented image-level and patch-level contrastive learning for global and local feature capture.
• Introduced a multiscale strategy and a feature similarity-based sampling method for enhanced adaptability and discriminability.

Main Results:

• Achieved a Dice score of 0.6296 after fine-tuning with only one annotated image.
• Outperformed existing self-supervised learning methods with identical model structures.
• Demonstrated the effectiveness of PSCL in microstructure segmentation.

Conclusions:

• PSCL is a highly effective self-supervised learning method for metallographic image segmentation.
• The method significantly reduces the need for extensive data annotation.
• PSCL offers a promising solution for microstructure identification in manufacturing processes.