Deep learning-based semantic segmentation of grain morphologies in ceramics

It is of great significance to study the distribution and quantity of various ceramic grains in ceramic materials for the research and improvement of their chemical and physical properties. There are many kinds of ceramic grain morphologies, which are different due to the phase structure or component composition. At present, the main way to distinguish different morphologies is to use the surface micrographs for artificial statistics, which can't effectively provide accurate information. This paper proposes a method of semantic segmentation of ceramic grains based on a fully convolution neural network. The method consists of three stages: building ceramic grain data set, training neural network, and testing network model. First, the grain morphologies of the ceramic surface were collected by a microscopic instrument, and the image was manually cut and marked with a marking software. Then, the cut image and the artificial label are input into the full convolution neural network for training. Finally, the test image is input into the trained network, and the segmented semantic image is output. The experimental results show that the semantic segmentation of crystal with a fully convolutional neural network can get good results, and the accuracy and speed are higher than that of artificial segmentation. Of most significance is the low cost of this method.