| Abstract [eng] |
This paper examines methods of surface metrology focusing on a class of optical methods. Application areas for these methods in the study of surface characteristics such as deformations are reviewed. Also, to study the images created by optical methods, machine learning methods for image recognition and classification are analyzed. The methodological part of this paper details the shadow moiré methodology for surface topography research. A strategy for generating images of shadow moiré interference fringes is described. To process these images, methodology of convolutional neural networks is chosen to solve the problem of classification. The parameters, algorithms and basics of AlexNet and LeNet architectures used by these networks are detailed. The software implementation used in the work is performed in Python programming language, using PlaidML framework and Keras library. In the exploratory part, the modification of the chosen convolutional neural network architecture is adapted to the task and the research of determining the optimal parameters is performed. Finally, a comparison and study of the effect of background image change on network results is performed. |
