Answers are grounded in this paper and cite their sources.
Rapid and reproducible segmentation of pore-like gaps in coal high-resolution transmission electron microscopy (HRTEM) images is valuable for quantitative microstructural analysis. We developed WT-UNet from a VGG-encoder U-Net by replacing conventional decoder-side detail convolution in all four feature-fusion stages with wavelet convolution (WTConv). The data set comprised 53 manually annotated original HRTEM images from 14 bituminous coals and 3896 retained pore regions after physical-scale harmonization. A coal-sample-level split assigned 37 images from ten coals to training, 8 images from two coals to validation, and 8 images from two other coals to independent testing. Training-only geometric and brightness transformations produced 222 training pairs. On the independent test set, WT-UNet achieved 92.73% pixel accuracy, 45.15% mean intersection over union (mIoU), 60.34% mean pixel accuracy (mPA), and 28.63% mean pore-class IoU (mIoUpore). The corresponding U-Net values were 92.58%, 44.07%, 58.96%, and 27.19%, while FCN-ResNet50 and DeepLabV3-ResNet50 achieved mIoUs of 42.88% and 43.30%. WT-UNet obtained the best aggregate performance among the four models and improved the IoU of all three pore-size categories relative to the matched U-Net. These results demonstrate that decoder-side wavelet feature processing improves multiscale pore representation and provides an efficient automatic tool for HRTEM pore-image analysis.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: