Stream pattern recognition of touching elements of grain mixtures

Abstract

The paper outlines the problem of stream classification of grain kernels, which involves machine learning methods in cases the images are touching. The applicability of the existing methods of contour separation under real industrial conditions is studied. As a result, a new efficient algorithm is introduced that enables stream separation of grain kernel images. The study has shown that morphological descriptors for restored contours can be used in feature vectors when solving the problem of image classification by means of a neural network with fully connected layers. A program solution is developed that enables classification of the elements of grain mixtures in industrial systems based on the use of 60 the most important morphological, colour and texture descriptors. The paper proposes using the statistical choice of descriptors and the automatic adjustment of the grain mixture flow while performing stream pattern recognition.