A dynamic-thresholding image segmentation algorithm for ceramic images based on Convolutional regression model

For the microscopic image of multiphase ceramics, the shape and size of the crystal grain in the traditional material electron microscopy images are manually marked by the researcher, its efficiency is poor. Because of the irregular texture on the crystal grain, it will cause large errors due to the large difference between the pixel value and the pixel attribute in the neighborhood by the traditional segmentation method. In this paper, a dynamic-thresholding image segmentation algorithm for multi-phase ceramic micro-images is presented. First of all, based on gradient image, it considers how to select the optimal threshold as the image label. Then, due to the sample is insufficient, it must meet the training requirements of the machine learning model, so it obtains pixel-level training samples from the labeled image, and use the modified convolution regression model to complete dynamic threshold automatic labeling to achieve image segmentation. From the experimental results, the algorithm accurately recognizes the weak edges of the ceramic image, while avoiding excessive noise, and the segmentation effect is significantly better than other algorithms. The accuracy of the model of automatically marking threshold is also higher than the other two models.