Deep-learning-based pyramid-transformer for localized porosity analysis of hot-press sintered ceramic paste

Area of Science:

• Materials Science
• Ceramic Engineering
• Image Analysis

Background:

• Scanning Electron Microscopy (SEM) is vital for ceramic microstructure analysis.
• Manual extraction of porosity from SEM images is labor-intensive and prone to error.
• Automated methods are needed to improve efficiency and accuracy in porosity quantification.

Purpose of the Study:

• To develop and validate PSTNet (Pyramid Segmentation Transformer Net) for automated grain and pore segmentation in ceramic SEM images.
• To accurately quantify porosity at ceramic grain boundaries.
• To enhance the efficiency and reliability of ceramic microstructure analysis.

Main Methods:

• PSTNet merges multi-scale feature maps for precise segmentation.
• The model was trained on Al2O3 and Y2O3 ceramic SEM datasets.
• A joint loss function incorporating segmentation penalty cross-entropy, smooth L1, and SSIM loss was employed.
• An improved multi-head attention mechanism was used for feature fusion in the decoder.

Main Results:

• PSTNet achieved a significant improvement in pixel accuracy (12.2% increase) and mean Intersection over Union (mIoU) (25.5% increase) over baseline methods.
• The model demonstrated low average relative errors of 6.9% for Y2O3 and 6.36% for Al2O3 datasets.
• Automated porosity measurement using PSTNet showed high accuracy and robustness.

Conclusions:

• PSTNet offers an effective automated solution for grain and pore segmentation in ceramic SEM images.
• The proposed method enhances the accuracy and efficiency of ceramic porosity quantification.
• This advancement facilitates more reliable analysis of ceramic material properties.