FastSAM-Based Automated Segmentation and Data Extraction for Pore Structures of Foamed Concrete

Reliable pore-structure recognition and descriptor extraction from foamed concrete micrographs remain challenging because pore walls are blurred, adjacent pores are often connected, and image statistics vary with observation scale. This study presents a FastSAM-DP workflow for automatic pore segmentation and pore-structure assessment. FastSAM generates pore instance masks, the Douglas–Peucker (DP) algorithm regularizes and simplifies contour geometry, and the workflow extracts the pore-size coefficient of variation (CV), circularity (Ci), number density (N), uniformity index (UI), large-pore area fraction (FL), and an image-derived composite pore-structure descriptor (PSQI). Configuration was selected using development data and assessed by source-filename-group held-out internal evaluation. In the complete 25-group/100-file held-out set, FastSAM-DP achieved an instance F1 of 0.732 (95% CI, 0.696–0.767), compared with 0.050 for fixed Otsu–Watershed; PSQI agreement was r = 0.828 (95% CI, 0.665–0.917), with MAE = 7.94 and RMSE = 11.10. In the post-audit 14-group/56-file non-overlap sensitivity subset, instance F1 was 0.752 (95% CI, 0.713–0.791) and the PSQI agreement was r = 0.838 (95% CI, 0.585–0.942), with MAE = 7.41 and RMSE = 10.50. Multiscale analysis included 130 images in 107 conservative image-field partitions at 15×, 20× and 40×, treated as independent descriptive strata. The workflow is therefore intended for batch pore-structure screening and same-magnification comparison within the present internal dataset.