INFORMATION AND MEASUREMENT SYSTEM FOR QUALITY CONTROL OF CERAMIC TILES BASED ON A KNOWLEDGE BASE FOR EMBEDDING IMAGES OF PRODUCTS OF THE CORRESPONDING GRADE

Abstract

The article considers the problem of automating the process of quality control and sorting of ceramic wall tiles. A critical analysis of currently implemented information and measurement systems based on expert selection of features of image fragments for quality control of the tile surface is conducted. The following features of such systems are identified: inflexibility and dependence on training data; high cost and duration of development (selection and justification of new features for each change in the technological process requires significant time and resources) low efficiency and high computational costs when calculating a large number of heterogeneous features of image segments; dependence on the scale of the image and defects (the informativeness of expertly selected features depends on the scale of the image and the size of the defects themselves); lack of a guarantee of optimality for manually selected features. It has been proven that this approach is inefficient, inflexible and resource-consuming, especially in the conditions of variable ceramic production. A new approach to creating an information-measuring quality control system based on the use of machine vision with a deep neural network and a vector database is proposed. Unlike existing solutions that use empirical functions and expert selection of image characteristics, the proposed system uses image embedding in a multidimensional semantic space using a pre-trained convolutional neural network. A structural and functional diagram of the analyzing and sorting process has been developed, which includes two main stages: preparation of a vector database with image embeddings of reference tiles and automation of sorting on the production line based on decision-making support for product quality by comparing image embeddings. The proposed solution avoids problems associated with the subjectivity of expert assessments and the high computational load of real-time control using previously implemented control systems.