METHODS AND DEVICES OF COMPUTER VISION IN QUANTITATIVE METALLOGRAPHY: A REVIEW OF FOREIGN INVENTIONS

The article presents a systematization of algorithmic solutions for automated analysis of metallographic images based on patent documentation from 2018 to 2026. The aim is to update information on foreign patents in the field of quantitative analysis of automated computer recognition processes for metallographic images. The paper considers the following: phase-segmentation and grain-boundary detection methods that employ deterministic digital-processing contours and modified loss functions to suppress topologically critical errors; approaches to augmenting training datasets via synthetic image generation using generative adversarial networks (GANs), followed by structure classification with support vector machines; and full-field methods for acquiring and stitching arrays of micrographs, including an entropy–information mutual-information criterion (entropy/joint entropy) and global optimization of FoV tile positions for stitching/stacking over multiple scanning cycles (in comparison with the conventional ASHLAR workflow E_NCC + minimum spanning tree). as well as data quality control loops (quality metric/threshold, rescanning), aggregation and filtering/rejection of frames based on quality metrics and hierarchical scanning schemes overview→target region, statistical post-processing of non-metallic inclusion detection results using extreme value statistics and the generalized Pareto distribution. It is shown that the key areas of patent analysis are associated with increasing the reproducibility of quantitative assessments due to formalizing segmentation topology, full-field representativeness and tail characteristics of defect distributions.