Deep-learning-based localized porosity analysis for laser-sintered Al2O3 ceramic paste

Abstract

We propose a deep-learning-based (DL-based) localized porosity analysis method to quantitatively map the surface porosity profile of the laser-sintered Al2O3 ceramic paste. A micro-pores detection and segmentation model was established by training the scanning electron microscopic (SEM) images of the laser-sintered Al2O3 ceramics using the Mask Region-based Convolutional Neural Network (Mask R–CNN). The obtained model was applied to automatically detect and segment the micro-pores out of the SEM images, and the surface porosity at the corresponding locations was calculated as the percentage of the pixels at the segmented pores within the SEM images. To improve the performance of the model, different training strategies were investigated to optimize the accuracy of micro-pores detection and segmentation. By comparing the AP50 values of the models trained by different strategies, the optimal model with an AP50 value of 0.894 was obtained after trained by a 101-layer residual neural network (ResNet) under supervised learning. To validate the developed models, a set of SEM images which is not used for the training processes has been applied to the surface porosity calculation. By calculating the surface porosity at the selected microscopic locations using the optimal model, the surface porosity profiles of the laser-sintered Al2O3 strips processed by different laser powers were quantitatively estimated. In addition, the surface porosity profiles calculated by the developed model were compared with the measurement results obtained from the existing image processing software to further evaluate the accuracy of the developed models.