High-precision grain size analysis of laser-sintered Al2O3 ceramics using a deep-learning-based ceramic grains detection neural network

Grain size is one of the most critical properties that significantly affect the performance of ceramic products. However, the analysis of grain size is still quite challenging due to their micro-scale features and irregular shapes. Herein, a deep-learning-based ceramic grains detection neural network (CGDNet) is developed for high-precision grain size analysis of the laser-sintered Al2O3 ceramics. The CGDNet is trained by the scanning electron microscopic (SEM) images of the laser-sintered Al2O3 ceramics using the Mask Region-Convolutional Neural Network (Mask R-CNN) algorithm. The obtained model is applied to automatically segment the ceramic grains out of the SEM images, and the grain size of the laser-sintered Al2O3 ceramics is estimated through calculating the area of the mask. To minimize the noise at the mask boundaries during grain segmentation, a low-resolution mask branch is developed to modify the regular Mask R-CNN. As a result, the performance of the CGDNet for grain segmentation is significantly improved compared to the regular Mask R-CNN, and the ceramic grains with irregular shapes can be precisely segmented. To demonstrate the superiority of CGDNet, the grain segmentation results of the developed model are compared with the commercial image processing software as well as other deep-learning-based image segmentation models, showing great improvement in the accuracy of ceramic grains detection and segmentation. By evaluating the average grain size at the selected microscopic locations using the CGDNet, the grain size distribution profile of the laser-sintered Al2O3 ceramics processed by different laser powers are quantitatively estimated.