Automatic detection and pixel-level quantification of surface microcracks in ceramics grinding: An exploration with Mask R-CNN and TransUNet

Microcrack damage on the grinding surfaces of engineering ceramics is inevitably incurred. To accurately evaluate the microcrack damage, an automatic detection and pixel-level quantification model based on the joint Mask R-CNN and TransUNet is developed. In addition, a joint training strategy is employed and the model is trained effectively on the image dataset of microcrack damage derived from the Si3N4 grinding, as captured by SEM. The Mask R-CNN demonstrates reliable automatic detection of microcracks, achieving an Average Precision of AP50 = 0.989 and AP75 = 0.864. Meanwhile, the TransUNet achieves fine segmentation of microcracks with complex characteristics, with an F1 score of 0.914 and an IoU value of 0.785. A skeleton-based quantification model of microcrack size is proposed, which delivers comprehensive and precise measurements of area, length, and notably, the width. The proposed quantification model provides a technical reference for the automatic evaluation of grinding surface quality.

Automated summary

An automatic detection and pixel-level quantification model based on joint Mask R-CNN and TransUNet is developed to accurately evaluate microcrack damage on the grinding surfaces of engineering ceramics. The model is effectively trained on actual micrograph image dataset using a joint training strategy. The proposed model achieves reliable automatic detection and fine segmentation of microcracks, and a skeleton-based quantification model is also proposed to provide comprehensive and precise measurements of microcrack size.